273 79 26MB
English Pages 261 Year 2021
IET COMPUTING SERIES 39
ReRAM-based Machine Learning
Other volumes in this series: Volume 1 Volume 3 Volume 4 Volume 5 Volume 6 Volume 7 Volume 8 Volume 9 Volume 13 Volume 14 Volume 15 Volume 18 Volume 20 Volume 23 Volume 24 Volume 25 Volume 26 Volume 29 Volume 32
Volume 33 Volume 34 Volume 35 Volume 40
Knowledge Discovery and Data Mining M.A. Bramer (Editor) Troubled IT Projects: Prevention and turnaround J.M. Smith UML for Systems Engineering: Watching the wheels, 2nd Edition J. Holt Intelligent Distributed Video Surveillance Systems S.A. Velastin and P. Remagnino (Editors) Trusted Computing C. Mitchell (Editor) SysML for Systems Engineering J. Holt and S. Perry Modelling Enterprise Architectures J. Holt and S. Perry Model-based Requirements Engineering J. Holt, S. Perry and M. Bownsword Trusted Platform Modules: Why, when and how to use them Ariel Segall Foundations for Model-based Systems Engineering: From patterns to models J. Holt, S. Perry and M. Bownsword Big Data and Software Defined Networks J. Taheri (Editor) Modeling and Simulation of Complex Communication M.A. Niazi (Editor) SysML for Systems Engineering: A Model-based approach, 3rd Edition J. Holt and S. Perry Data as Infrastructure for Smart Cities L. Suzuki and A. Finkelstein Ultrascale Computing Systems J. Carretero, E. Jeannot and A. Zomaya Big Data-enabled Internet of Things M. Khan, S. Khan and A. Zomaya (Editors) Handbook of Mathematical Models for Languages and Computation A. Meduna, P. Horáˇcek and M. Tomko Blockchains for Network Security: Principles, technologies and applications H. Huang, L. Wang, Y. Wu and K.R. Choo (Editors) Network Classification for Traffic Management: Anomaly detection, feature selection, clustering and classification Zahir Tari, Adil Fahad, Abdulmohsen Almalawi and Xun Yi Edge Computing: Models, technologies and applications J. Taheri and S. Deng (Editors) AI for Emerging Verticals: Human-robot computing, sensing and networking M.Z. Shakir and N. Ramzan (Editors) Big Data Recommender Systems Volumes 1 and 2 Osman Khalid, Samee U. Khan and Albert Y. Zomaya (Editors) E-learning Methodologies: Fundamentals, technologies and applications M. Goyal, R. Krishnamurthi and D. Yadav (Editors)
ReRAM-based Machine Learning Hao Yu, Leibin Ni and Sai Manoj Pudukotai Dinakarrao
The Institution of Engineering and Technology
Published by The Institution of Engineering and Technology, London, United Kingdom The Institution of Engineering and Technology is registered as a Charity in England & Wales (no. 211014) and Scotland (no. SC038698). © The Institution of Engineering and Technology 2021 First published 2021 This publication is copyright under the Berne Convention and the Universal Copyright Convention. All rights reserved. Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may be reproduced, stored or transmitted, in any form or by any means, only with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms of licences issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to the publisher at the undermentioned address: The Institution of Engineering and Technology Michael Faraday House Six Hills Way, Stevenage Herts, SG1 2AY, United Kingdom www.theiet.org While the authors and publisher believe that the information and guidance given in this work are correct, all parties must rely upon their own skill and judgement when making use of them. Neither the authors nor publisher assumes any liability to anyone for any loss or damage caused by any error or omission in the work, whether such an error or omission is the result of negligence or any other cause. Any and all such liability is disclaimed. The moral rights of the authors to be identified as authors of this work have been asserted by them in accordance with the Copyright, Designs and Patents Act 1988.
British Library Cataloguing in Publication Data A catalogue record for this product is available from the British Library
ISBN 978-1-83953-081-4 (hardback) ISBN 978-1-83953-082-1 (PDF)
Typeset in India by MPS Limited Printed in the UK by CPI Group (UK) Ltd, Croydon
Contents
Acronyms Preface About the authors Part I Introduction
ix xi xv 1
1 Introduction 1.1 Introduction 1.1.1 Memory wall and powerwall 1.1.2 Semiconductor memory 1.1.3 Nonvolatile IMC architecture 1.2 Challenges and contributions 1.3 Book organization
3 3 3 5 11 14 16
2 The need of in-memory computing 2.1 Introduction 2.2 Neuromorphic computing devices 2.2.1 Resistive random-access memory 2.2.2 Spin-transfer-torque magnetic random-access memory 2.2.3 Phase change memory 2.3 Characteristics of NVM devices for neuromorphic computing 2.4 IMC architectures for machine learning 2.4.1 Operating principles of IMC architectures 2.4.2 Analog and digitized fashion of IMC 2.4.3 Analog IMC 2.4.4 Digitized IMC 2.4.5 Literature review of IMC 2.5 Analysis of IMC architectures
19 19 20 21 22 23 24 25 26 27 29 34 34 40
3 The background of ReRAM devices 3.1 ReRAM device and SPICE model 3.1.1 Drift-type ReRAM device 3.1.2 Diffusive-type ReRAM device 3.2 ReRAM-crossbar structure 3.2.1 Analog and digitized ReRAM crossbar 3.2.2 Connection of ReRAM crossbar
45 45 45 52 54 55 57
vi ReRAM-based machine learning 3.3 ReRAM-based oscillator 3.4 Write-in scheme for multibit ReRAM storage 3.4.1 ReRAM data storage 3.4.2 Multi-threshold resistance for data storage 3.4.3 Write and read 3.4.4 Validation 3.4.5 Encoding and 3-bit storage 3.5 Logic functional units with ReRAM 3.5.1 OR gate 3.5.2 AND gate 3.6 ReRAM for logic operations 3.6.1 Simulation settings 3.6.2 ReRAM-based circuits 3.6.3 ReRAM as a computational unit-cum-memory
59 61 61 62 63 65 67 70 70 70 71 72 73 74
Part II Machine learning accelerators
77
4 The background of machine learning algorithms 4.1 SVM-based machine learning 4.2 Single-layer feedforward neural network-based machine learning 4.2.1 Single-layer feedforward network 4.2.2 L2-norm-gradient-based learning 4.3 DCNN-based machine learning 4.3.1 Deep learning for multilayer neural network 4.3.2 Convolutional neural network 4.3.3 Binary convolutional neural network 4.4 TNN-based machine learning 4.4.1 Tensor-train decomposition and compression 4.4.2 Tensor-train-based neural network 4.4.3 Training TNN
79 79 80 80 84 87 87 87 88 93 93 94 96
5 XIMA: the in-ReRAM machine learning architecture 5.1 ReRAM network-based ML operations 5.1.1 ReRAM-crossbar network 5.1.2 Coupled ReRAM oscillator network 5.2 ReRAM network-based in-memory ML accelerator 5.2.1 Distributed ReRAM-crossbar in-memory architecture 5.2.2 3D XIMA
99 99 99 106 108 109 111
6 The mapping of machine learning algorithms on XIMA 6.1 Machine learning algorithms on XIMA 6.1.1 SLFN-based learning and inference acceleration 6.1.2 BCNN-based inference acceleration on passive array 6.1.3 BCNN-based inference acceleration on 1S1R array
115 115 115 117 121
Contents L2-norm gradient-based learning and inference acceleration 6.1.5 Experimental evaluation of machine learning algorithms on XIMA architecture 6.2 Machine learning algorithms on 3D XIMA 6.2.1 On-chip design for SLFN 6.2.2 On-chip design for TNNs 6.2.3 Experimental evaluation of machine learning algorithms on 3D CMOS-ReRAM
vii
6.1.4
122 126 141 141 145 151
Part III Case studies
165
7 Large-scale case study: accelerator for ResNet 7.1 Introduction 7.2 Deep neural network with quantization 7.2.1 Basics of ResNet 7.2.2 Quantized convolution and residual block 7.2.3 Quantized BN 7.2.4 Quantized activation function and pooling 7.2.5 Quantized deep neural network overview 7.2.6 Training strategy 7.3 Device for in-memory computing 7.3.1 ReRAM crossbar 7.3.2 Customized DAC and ADC circuits 7.3.3 In-memory computing architecture 7.4 Quantized ResNet on ReRAM crossbar 7.4.1 Mapping strategy 7.4.2 Overall architecture 7.5 Experiment result 7.5.1 Experiment settings 7.5.2 Device simulations 7.5.3 Accuracy analysis 7.5.4 Performance analysis
167 167 168 168 170 172 172 173 173 174 174 176 176 177 177 178 180 180 181 182 185
8 Large-scale case study: accelerator for compressive sensing 8.1 Introduction 8.2 Background 8.2.1 Compressive sensing and isometric distortion 8.2.2 Optimized near-isometric embedding 8.3 Boolean embedding for signal acquisition front end 8.3.1 CMOS-based Boolean embedding circuit 8.3.2 ReRAM crossbar-based Boolean embedding circuit 8.3.3 Problem formulation 8.4 IH algorithm 8.4.1 Orthogonal rotation
189 189 192 192 192 194 194 195 197 197 198
viii
ReRAM-based machine learning 8.4.2 Quantization 8.4.3 Overall optimization algorithm 8.5 Row generation algorithm 8.5.1 Elimination of norm equality constraint 8.5.2 Convex relaxation of orthogonal constraint 8.5.3 Overall optimization algorithm 8.6 Numerical results 8.6.1 Experiment setup 8.6.2 IH algorithm on high-D ECG signals 8.6.3 Row generation algorithm on low-D image patches 8.6.4 Hardware performance evaluation
199 199 200 200 201 202 203 203 204 207 210
9 Conclusions: wrap-up, open questions and challenges 9.1 Conclusion 9.2 Future work
215 215 216
References Index
217 237
Acronyms
ANN BCNN BNN CNN DNN DRAM FPGA HDD IMC ML NDC NVM ODD PCM PIM ReRAM ResNet SLFN SRAM STDP STT-MTJ STT-RAM TNN TSV XIMA 1S1R 1T1R
artificial neural network binary convolutional neural network bitwise neural network convolutional neural network deep neural network dynamic random-access memory field programmable gate array hard disk drive in-memory computing machine learning near-data computing nonvolatile memory optical disk drive phase change memory processing-in-memory resistive random-access memory residual network single-layer feedforward neural network static random-access memory spike timing-dependent plasticity spin transfer torque magnetic tunnel junction spin transfer torque RAM tensor neural network through-silicon via crossbar in-memory architecture one selector one ReRAM one transistor one ReRAM
This page intentionally left blank
Preface
With the emergence of IoT and handheld devices, the amount of data procured in the data centers have reached nearly exa-scale. Processing such large amounts of data through traditional computing techniques is inefficient due to latency and inefficient resource usage. With the introduction in machine learning (ML) and success in multiple application, ML has been adopted for big data processing. Despite advancements in terms of processing through ML, transferring and communicating the data to and from the memory and storage units is seen as one of the major bottlenecks. In-Memory Computing (IMC) is seen as a panacea to overcome the challenges of traditional Von-Neumann architectures. For efficient IMC, frameworks such as Hadoop and MapReduce frameworks have been introduced. Such frameworks explore the temporal locality to process such large amounts of data. Despite efficient compared to traditional computing paradigms, existing IMC paradigms are inefficient in terms of power consumption and latency requirements, especially when employed in data centers and other cloud computing platforms. To address the challenges with CMOS-based IMC, emerging non-volatile memory (NVM) devices are researched in both academia and industry. Such NVM-based IMC architectures are observed to overcome the saturation of Moore’s law challenge because of its low power, area and high-density embedded storage benefits. Multiple devices emerging from different materials including resistive random-access memory (ReRAM), phase-change random access memory (PCRAM), magnetic RAM (MRAM), Ferroelectric RAM (FeRAM) and NOR Flash along with traditional DRAMs are introduced and researched. Among multiple emerging NVM devices, ReRAM is the most promising among multiple devices due to its potential of multi-bit storage and compatability with CMOS technology. Such exploration is well-supported by companies such as HP and IBM. Hence, we believe the ReRAM-based IMC will be commercialized with IoT products in the next few years. This book first introduces existing architectures on IMC such as processingin-memory (PIM), near-data processing (NDP) and near-data computation (NDC) architectures ranging from general programmable many-core systems to reconfigurable arrays and custom-designed accelerators. A comprehensive analysis on various IMC architectures is discussed and compared. Further, with the emergence of memory devices that support IMC, the authors introduce the ReRAM along with the device modeling and logic circuit design needed for ML operations. ReRAM has been widely deployed for IMC architectures due to minimized leakage power, reduced power consumption and smaller hardware footprint. The authors introduce the design of a crossbar structure with the aid of ReRAM to facilitate matrix multiplications and
xii
ReRAM-based machine learning
convolutions that are computationally intensive operations in ML applications. More importantly, they present a true in-memory logic-integration architecture without utilizing I/Os that leads to faster data access and higher throughput with low power. They also present multiple ReRAM-based IMC architectures that can perform other computations of ML and data-intensive applications such as calculation of gradients and L2-norms. Deep learning architectures are heavily growing in terms of depth and leading to an exponential increase in computational complexity. As a case study, the authors discuss how the ReRAM-based IMC can aid in accelerating computations, eventually leading to an increased throughput with high energy efficiency and smaller form factor for a large-scale deep learning network such as residual network (ResNet) and compressive sensing computations. As the number of elements on a single chip or system are limited, to further enhance the throughputs of data-intensive computations, a distributed computing paradigm has been widely adopted. In this book, the authors introduce techniques to perform distributed computing using IMC accelerators. The emergence of 3D integration has facilitated multiple chips stacked on top of each other, with each chip functioning independently. The authors introduce a 3D CMOS-ReRAM architecture, where ReRAM devices are connected through through-silicon-vias (TSVs) with toplayer wordlines and bottom-layer bitlines. All other CMOS logics are implemented in the bottom layer. They also introduce a multilayer 3D architecture, where multiple ReRAM layers are designed as data buffers. This book introduces the design of hardware for ML and data-intensive computations. It has been shown by the research community that the mapping of computations is a critical aspect without which the benefits of IMC or hardware accelerator designs cannot be yielded. Thus, to make the design complete, the authors introduce strategies to map ML designs onto hardware accelerators. They also provide data analytics mapping on ReRAM-based IMC architecture for multiple data-intensive applications. The learning and inference procedures of single-layer feedforward neural network (SLFN) have been optimized and partially mapped on a passive binary ReRAM crossbar. In addition, the authors map a binary convolutional neural network (BCNN) on both passive array and One Selector One ReRAM array with different mapping schemes. In addition to traditional 2D integration, this book also introduces how mapping can efficiently be performed for emerging 3D integration-based IMC architectures. This book is highly motivated by the rapid advancements observed in the ML and hardware design communities, and the lack of a book that provides insights and latest advancements in both these fields, as well as the wide adoption of ML algorithms in a plethora of applications and ever-increasing industrial requirements for enhancing the performance of ML applications. This book can serve as a bridge between researchers in the computing domain (algorithm designers for ML) and hardware designers. The first chapter introduces the need and motivation to deploy ML algorithms for data-intensive computations. This also introduces the challenges and need for hardware for ML processing. In the second chapter, we present emerging memory devices that can support IMC and the need for IMC is introduced in this chapter. Further, a comprehensive comparison and analysis of different IMC architectures is presented. The architectures include using traditional DRAM to emerging devices
Preface
xiii
such as ReRAM and memristors. In the third chapter, modeling of ReRAM devices is presented followed by the design of crossbar structures that are primarily used for matrix multiplication in ML applications. In addition, design of multibit storage for ReRAMs and other logic functional units with ReRAMs is presented. The fourth chapter provides the basic information on ML algorithms, computational complexity and the types of operations involved in the ML algorithms, including least-squares method (SVM, ELM), and deep learning (CNN, TNN). Trained quantization (binary, ternary and eight-bit, etc.) will be addressed in this chapter. Additionally, the training mechanisms for deep CNNs and TNNs are focused in this chapter, as they are the time-consuming operations and the performance depends on the training efficiency. The fifth chapter introduces multiple strategies to map the ML designs on to hardware accelerators. Designing of ReRAM crossbar-based IMC architectures to accelerate ML applications is presented. Further, to accommodate deep DNNs and CNNs, a 3D stacking-based CMOS-ReRAM architecture is introduced. The main focus of this chapter is on the architectural aspects and circuity to support the dataintensive computations with high speed. In Chapter 6, we introduce techniques to map the ML applications for ReRAM and the major challenges and the scope for improvement. The mapping techniques for different ML algorithms for different architectures such as ReRAM cross with traditional 2D and 3D architectures are illustrated. Case studies for ReRAM-based ML acceleration for different scenarios are presented. Chapter 7 presents ResNet under different scenarios such as DNN with quantization and ReRAM-based acceleration for ResNet. In addition to the ResNet, this book presents ReRAM-based accelerator for compressive sensing application in Chapter 8. For this part, different scenarios such as iterative heuristic algorithm and row generation algorithms are presented. All the case study chapters are presented with substantial amount of experimental results and analysis that aid in replicating the experiments and also prove the efficiency of ReRAM accelerators and its prominence in the design of future hardware accelerators. The last chapter summarizes the ReRAM accelerators and its impact on the ML applications. Furthermore, future directions with ReRAM devices to further enhance the ML computations are presented. The authors of this book would like to thank the colleagues at Nanyang Technological University, especially members of VIRTUS lab, Hardware architecture and artificial intelligence lab at George Mason University and the affiliated students. We would like also to express my deepest appreciation to the faculty members, Prof. Anupam Chattopadhyay, Prof. Chang Chip Hong, Prof. Shi Guoyong, Prof. Huang Guangbin of Nanyang Technological University and Prof. J. Joshua Yang and Prof. Qiangfei Xia of the University of Massachusetts Amherst for their support. The authors also would like to thank Prof. Amlan Ganguly and Mr. Purab Sutradhar of Rochester Institute of Technology. The authors would also like to thank Dr. Nima Taherinejad and Prof. Axel Jantsch from Vienna University of Technology (TU Wien), Austria.
This page intentionally left blank
About the authors
Hao Yu is a professor in the School of Microelectronics at Southern University of Science and Technology (SUSTech), China. His main research interests cover energy-efficient IC chip design and mmwave IC design. He is a senior member of IEEE and a member of ACM. He has written several books and holds 20 granted patents. He is a distinguished lecturer of IEEE Circuits and Systems and associate editor of Elsevier Integration, the VLSI Journal, Elsevier Microelectronics Journal, Nature Scientific Reports, ACM Transactions on Embedded Computing Systems and IEEE Transactions on Biomedical Circuits and Systems. He is also a technical program committee member of several IC conferences, including IEEE CICC, BioCAS, A-SSCC, ACM DAC, DATE and ICCAD. He obtained his Ph.D. degree from the EE department at UCLA, USA. Leibin Ni is a Principle engineer at Huawei Technologies, Shenzhen, China. His research interests include emerging nonvolatile memory platforms, computing inmemory architecture, machine learning applications and low power designs. He is a member of IEEE. He received his Ph.D. from the Nanyang Technological University, Singapore. Sai Manoj Pudukotai Dinakarrao is an assistant professor in the Department of Electrical and Computer Engineering at George Mason University (GMU), USA. His current research interests include hardware security, adversarial machine learning, Internet of things networks, deep learning in resource-constrained environments, inmemory computing, accelerator design, algorithms, design of self-aware many-core microprocessors and resource management in many-core microprocessors. He is a member of IEEE and ACM. He served as a guest editor to IEEE Design and Test Magazine and reviewer for multiple IEEE and ACM journals. Also, he is a technical program committee member of several CAD conferences, including ACM DAC, DATE, ICCAD, ASP-DAC, ESWEEK and many more. He received a Ph.D. degree in Electrical and Electronic Engineering from the Nanyang Technological University, Singapore.
This page intentionally left blank
Part I
Introduction
This page intentionally left blank
Chapter 1
Introduction
1.1 Introduction In the last two decades, computing paradigms have experienced a tremendous drift with nearly exascale (1018 bytes) amount of data being amassed by business units and organizations. The requirements to analyze and respond to such large amounts of data have led to the adoption of machine learning (ML) methods in a wide range of applications ranging from autonomous driving to smart homes. One of the major challenges to process such large amounts of data is the fetching of data from the memory and writing it back without experiencing the well-known memory-wall bottleneck. To address such concerns, in-memory computing (IMC) and supporting frameworks such as Hadoop and MapReduce have been introduced. This has shown to be highly beneficial as these methods address the challenge of a lack of temporal locality that exists in traditional computing techniques. Undoubtedly, these techniques have helped but the challenges of power consumption and processing delays are still unsolved concerns, especially in macro systems and data centers. Computing-in-memory methods based on emerging nonvolatile devices are popular in both academia and industry. Many researchers believe that this architecture will be an opportunity to break Moore’s law because of its ultra-low power and high-density embedded storage. Various devices including resistive random-access memory (ReRAM), phase-change random-access memory (PCRAM), Magentic RAM (MRAM), Ferroelectric RAM (FeRAM) and NOR Flash have been discussed. ReRAM is the most promising among all these devices due to its potentiality for multilevel resistance and compatibility with CMOS technology. Well-known companies such as IBM and HP have invested in this field, and we believe that this ReRAM-based IMC will be commercialized with Internet-of-things (IoT) products in the next 2–3 years. This chapter introduces the need and motivation to deploy ML algorithms for data-intensive computations.
1.1.1 Memory wall and powerwall Information technology by new social media such as Google and Facebook is generating a wealth of data at exa-level that is previously unimaginable. For instance, WalMart processes one million transactions per hour, and from these transactions, they collect
4 ReRAM-based machine learning 2.5 petabytes of information [1], and Facebook users upload 100 terabytes of data and one billion pieces of content every day [2]. In-depth data analytics may become the driving force for the future economic growth in a digital data-oriented society. The main challenge stems from the existing hardware infrastructure built upon the conventional computer architecture and CMOS technology. Moore’s law scaling has become slower when compared to the growth of data. There are many recent largescale research programs such as exascale computing (NSF/DARPA-USA) and Human Brain (European Commission), which are, however, still limited at this singularity point because of the memory bottleneck. To process and gain insights from such massive collected data, parallel and scalable ML-based solutions that can scale with data are needed. In response, several scalable ML algorithms along with parallel frameworks have emerged: Hadoop [3], Spark [4], Flink [5], [6], Tez [7], Google Dataflow [8], 0xdata H2O [9] and Petuum [10]. Undoubtedly, these frameworks pushed the processing capabilities beyond traditional approaches, but executing these emerging analytic solutions requires a significant amount of computational resources such as power and appropriate processing architectures. Thus, the data centers started to expand by equipping more computational nodes to satisfy the demands of increasing volumes of data [11–14] and the complexity of ML algorithms. While demand for ML-based computational resources continues to grow with the size of data, the semiconductor industry has reached the scaling limits and is no longer able to reduce power consumption in new chips. Current server designs, based on commodity homogeneous processors, ceased to be efficient in terms of performance/watt to process such ML-based data-intensive computations [15–22]. During big-data analytics, the computing systems in data centers or nodes need to process the stored data with intensive memory accesses. Memory access is seen as one of the pivotal obstacles that limit the performance and throughput of the current architectures and data centers. Intensive memory access is seen across the emerging applications including web searching, data mining and also ML applications. The microprocessor needs to process the stored data with the intensive memory access. However, the state-of-the-art data storage and processing hardware still have a bandwidth-wall problem with large I/O congestion, but also powerwall with large static power in advanced CMOS technology. As a result, a design of scalable energy-efficient big-data analytic hardware is highly required. The following research activities [23,24] attempted to alleviate the memory bottleneck problem for big-data storage and processing in the future: ●
First, as a promising technology for big-data-oriented computing, the emerging nonvolatile IMC shows a great potential as a high-throughput, energy-efficient solution. The application-specific accelerators can be developed within memory. In this case, before readout by processing unit, data can be preprocessed so that data migration is minimized and the bandwidth wall can be relieved. Moreover, the nonvolatile memory (NVM) devices hold information without using charge such that the leakage current can be significantly reduced with the powerwall relieved as well.
Introduction ●
●
5
Second, sparse-represented data by compressive sensing is an effective approach to reduce data size by projecting data from high-dimensional space to lowdimensional subspace with essential feature preserved. Data stored in the memory after compressive sensing can be recovered or directly processed. This procedure reduces data complexity in data storage as well as in data analytics with a further significant power reduction and bandwidth improvement. Third, data analytics by ML accelerator can be developed for fast data analytics. Latest ML algorithm on-chip can accelerate the learning process with a potential to realize online training for efficient inference, which is important for applications in autonomous vehicles and unmanned aerial vehicles.
1.1.2 Semiconductor memory As seen in the previous section, having a memory element close to the processing units can aid plethora of challenges. There exist numerous kinds of memory which are different in terms of their functionality, physical characteristics as well as the benefits they provide. We first review the memory utilized in computer architecture, which will lay basis for the architectural design described in the later chapters.
1.1.2.1 Memory technologies Traditional semiconductor memories refer to the silicon-based transistor devices (CMOS, BTJ, FINFET and so on) in computer architectures. These memories are applied to store computation instructions as well as data temporarily or permanently. Compared to storage media such as hard disk drive (HDD) and optical disk drive (ODD) with predetermined order of data access due to the limitations of mechanical drive, the semiconductor memories possess the property of random access. It takes almost an identical time to access any data regardless of the data location. One can further classify all the semiconductor memories into two classes: volatile memory and NVM. In volatile memory, data are stored as electrical signals such as voltage, and it requires uninterrupted power supply. Once the power supply stops, the device is turned off and data are lost. Currently, the most commonly used volatile memories are static random-access memory (SRAM) [25–27] and dynamic randomaccess memory (DRAM), whose data are indicated by the electrical voltage levels. On the contrary, NVM is able to retain the data even when the device is turned off, as its data are mostly preserved by nonelectrical states. For instance, bits in programmable read-only memory (PROM) are denoted by whether the fuses of individual memory cells are burned. Figure 1.1 briefly introduces the characteristics of the volatile SRAM/DRAM, traditional nonvolatile memories (NVMs) and also emerging NVMs. The memory elements are arranged in terms of the size and access time (the left most elements are present in small quantity/size, but have faster access to the processor, whereas towards the right the elements are in large quantity and have large access time). Three prominent memory technologies are the SRAM, DRAM and flash memory, whose details are presented in the following.
6 ReRAM-based machine learning Volatile
Emerging NVM
SRAM
Traditional NVM Memristor STT-MTJ
Magnetic disk
PCM DRAM
100
101
102
103
Register L1L2L3 Main cache memory
Magnetic tape
NAND
104
105
SSD
106
107
108
109
1010
1011
HDD
Figure 1.1 Hierarchy of different memories in a computer architecture WL VDD M2
M4 M6
M5
Q
Q BL
BL M1
M3
Figure 1.2 A 6T SRAM cell structure with leakage paths in standby state. The BLs are precharged high and assume the stored data is “1” at Q
Static random-access memory A typical CMOS-SRAM cell consists six transistors, shown in Figure 1.2. The flipflop formed by M1 to M4 holds the stored bit. The term static is derived from the fact that the cell does not need to be refreshed like dynamic RAM, and the data can be retained as long as the power is supplied. The M5 and M6, connected with WL and two bitlines (BLs), are used as access transistors to select target cells.
Introduction
7
There are three operation states for an SRAM cell, write, read and stand by. To a write operation, the value to be written needs to be applied on both BLs, namely BL and BL, in a complementary manner. Assume we wish to write “0” to the cell, i.e., Q to be “0” and Q to be “1,” the BL is driven low and BL high. Once the M5 and M6 are turned on by setting WL “1,” the BL drivers will override the previous stored value. In order to easily override the previous state in the self-reinforced flip-flop, the BL drivers are required to be designed stronger than the transistors in flip-flop. For a read operation, both BLs are precharged high before the start of the read cycle, and the turning on of the wordline (WL) signifies the start of read operation. Because of the opposite voltages at Q and Q, one of the BLs will be pulled down by the cell, and the discharging of one of the BLs is then detected by the BL sense amplifier. In voltage mode sensing scheme, the sign of BL voltage difference V (VBL minus VBL ) determines the value of stored bit. A V in tens of millivolts is significant enough to efficiently distinguish which BL is being discharged. For example, assume the stored bit is “1” at Q and “0” at Q, and once the WL is asserted, the BL will discharge towards “0”; when a positive V of tens of millivolts is gained, the latch-based sense amplifier will amplify the small voltage difference with positive feedback, and finally output logic “1” as result. When the WL is connected to ground, turning off the two access transistors, the cell is in the standby state. During the standby state, the two cross-coupled inverters will reinforce each other due to the positive feedback, the value is preserved as long as the power is supplied. One prominent problem regarding SRAM in standby state is severe subthreshold leakage. Subthreshold leakage is the drain-source current of a transistor when the gate-source voltage is less than the threshold voltage. The subthreshold current depends exponentially on threshold voltage, which results in large subthreshold current in deep sub-micron regime. Figure 1.2 shows three leakage paths in one SRAM cell, assuming the stored bit is “1” at Q. Note that the BL and BL are always precharged to VDD to facilitate future read operation. Regardless of the stored value, there always will be three transistors consuming leakage power. Compared to other memory technologies, SRAM is able to provide the fastest access speed, but the advantage comes as a trade-off against density and power. As one SRAM cell requires silicon area for six transistors, SRAM has very limited density and hence is more expensive than other memory technologies. In addition, it is very power consuming due to the leakage problem at standby state. Therefore, SRAM serves best in the applications where high performance is the main concern and the capacity is not significant, namely the caches for processors.
Dynamic random-access memory The philosophy behind DRAM is simplicity. Unlike SRAM, where one cell is composed of six transistors, each individual DRAM cell consists only one capacitor and one access transistor. The data “0” or “1” is represented by whether the capacitor is fully charged or discharged. However, the electrical charge on the capacitor will gradually leak away, and after a period, the voltage on the capacitor is so low for the sense amplifier to differentiate between “1” and “0”. Therefore, unlike SRAM
8 ReRAM-based machine learning that the data can be retained as long as the power is supplied, the retention time for DRAM is finite and all DRAM cells need to be read out and written back periodically to ensure data integrity. Typically, the cells are refreshed once every 32 or 64 ms. This process is known as refresh , and this is how the name of dynamic RAM is derived. Figure 1.3 shows the circuit diagram of such 1T1C structure DRAM cell. To write a DRAM cell, the BL is first set high or low based on the value to write. After the access transistor is turned on by asserting WL, the capacitor in the selected cell is charged to “1” or discharged to “0.” Because the access takes place through an NMOS transistor, there exists a Vth drop during the write “1.” In order to prevent this Vth drop and maintain a long refresh period, the WLs driving voltage is usually boosted to VPP = VDD + Vth . To read a DRAM cell, the BL is precharged to VDD /2 and then the WL is enabled. Due to the charge sharing, the BL voltage will slightly decrease or increase depending on the voltage that the capacitor was previously charged to, i.e., VDD or 0. If it is previously charged, the charge sharing will slightly boost the BL voltage; otherwise, some charge will be distributed from BL to cell capacitor. In both cases, the voltage of storage capacitor will be changed after read operation; thus, the read operation is called destructive, and an instant write back is required. A slight voltage change on the BL can be calculated by V = ±
VDD Ccell 2 Ccell + Cbitline
(1.1)
The sign of V depends on the state of storage capacitor. In modern DRAM devices, the capacitance of a storage capacitor is far smaller than the capacitance of the BL. Typically, the capacitance of a storage capacitor is one-tenth of the capacitance of the long BL that is connected to hundreds or thousands of other cells. The relative capacitance values create the scenario that when the small charge contained in a storage capacitor is placed on the BL, the resulting voltage on the BL is small and difficult to measure in an absolute sense. In DRAM devices, the voltage sensing problem is resolved through the use of a differential sense amplifier that compares the voltage of the BL to a reference voltage.
WL
BL
Ccell
Figure 1.3 The circuit diagram of 1T1C DRAM cell structure
Introduction
9
The use of differential sense amplifier, in turn, introduces some requirements on the DRAM array structure. Particularly, instead of a single BL, a pair of BLs needs to be used to sense the voltage value contained in any DRAM cell. In addition, in order to ensure that the voltage and capacitance values on the pair of BLs are closely matched, the BLs must be closely matched in terms of path lengths and the number of cells attached. The above requirements lead to two distinctly different array structures: open BL structures and folded BL structures.
Flash NVM Flash memory is the most widely used NVM technology today. The key device in this prevailing memory is floating gate (FG) transistors. A figure of cross section of an FG transistor is shown in Figure 1.4. Unlike a metal oxide silicon field effect transistor (MOSFET) transistor, an additional floating gate is added between the control gate and channel. Isolated by oxide layers, floating gate is able to trap charges and keep them for years. Therefore, the FG transistor encodes data based on whether electrons are trapped and is able to retain data without power. That is where “nonvolatile” is derived from. The principle of read operation can be described as follows. When no charges are trapped in floating gate, the FG transistor has a threshold voltage of Vth0 ; when negative charges are trapped, they attract positive charges of control gate; thus, higher control gate voltage is required to turn on the channel, which produces a higher threshold voltage Vth1 . By applying intermediate control gate voltage that is between Vth0 and Vth1 , and measuring current, the state of device can be known. The write operation of FG transistor involves injecting or pulling electrons across the oxide barrier. There are two ways to achieve this, quantum tunneling or hot electron injection. In quantum tunneling scenario, high voltage is applied on control gate, quantum tunneling will take place between the floating gate and channel, and electrons can travel across the oxide barrier. For hot electron injection scenario, electrons are accelerated under high electrical field in the channel till its energy is high enough to penetrate the oxide layer. Note that electrons with high energy will damage the oxide lattice, and such damage will accumulate and lead to a limited write cycles, which is typically around 105 cycles.
Control gate Oxide
Floating gate
n+
n+ p-substrate
Figure 1.4 The cross section of an FG transistor
10
ReRAM-based machine learning
There are two common layouts for flash memory, shown in Figure 1.5, NAND flash memory with FG transistors in series and NOR flash memory with FG transistors in parallel. The names NAND and NOR are derived from the fact that their connection fashion in series or parallel resemble a NAND gate or a NOR gate. NAND layout has the density advantage over NOR layout because each row only has one ground connection, and is thus widely used for external storage. NOR layout has lower latency and is thus widely used in embedded systems, where high performance is required. Figure 1.5 also shows how the read of NAND and NOR flash memory can be achieved. The relationship between the different applied voltage magnitude is shown as follows: VOFF < Vth0 < VINT < Vth1 < VON |VHIGH |
(1.2)
1.1.2.2 Nanoscale limitations With scaling down of the technology towards nanolevel, there are some challenges and limitations in the traditional memory architecture. Such challenges are caused by physical rules and is summarized from two perspectives. On the one hand, there are functional failures due to the process variation that transistors may have mismatch problems. On the other hand, the thermal runaway failures are caused by positive feedback in an SRAM or a DRAM cell [28]. As a result, NVM is considered as a potential solution for data storage with the scaling technology.
Bitline VOFF
VOFF
VOFF
VINT
VOFF
VOFF
VOFF
VOFF
VON
VON
VON
NOR flash memory
Bitline
VON
VON
VON
VINT
VON
NAND flash memory
Figure 1.5 Two common layouts for flash memory: NOR flash memory and NAND flash memory
Introduction
11
1.1.3 Nonvolatile IMC architecture In the conventional computing architecture, all the data are stored in memory. It is separated from the general processors so that I/O connections are required. As a result, during the data processing, all data need to be read out to the external processor and written back afterwards. However, in data-oriented applications such as ML acceleration, this architecture will incur significant I/O congestions and hence greatly degrade the overall performance [27,29–32]. In addition, significant static power will be consumed because all the data need to be held, even though it does not need to be used at the moment [26,28,33–39]. Theoretically, it is feasible to overcome the bandwidth issue by adding more I/O pins or operating them at higher frequency. Practically, however, the I/O frequency is limited by the signal propagation delay and signal integrity issues, and I/O number is limited by the packaging technology, and thus the bandwidth is nearly saturated. Besides improving the bandwidth of memory, reducing the volume of data communicating between processor and memory is another feasible option to enhance the performance. Conventionally, the processor only reads raw data from the main memory. If the memory can perform some operations before sending data, the I/O communication requirement can be significantly reduced [40–42]. For instance, if one needs to add eight numbers, one has to load all the eight values into the processor in the conventional way. However, if the addition of two numbers is performed as an in-memory logic operation, one can preprocess the adding inside memory and only four numbers need to be read out. To perform IMC, implementing logic operations inside the memory is required so that the preprocessing can be done. Such an architecture is called logic-in-memory architecture. In general, the requirement of the ideal logic-in-memory architecture is summarized in Figure 1.6. A big NVM sea is connected with thousands of small accelerator cores through high bandwidth and energy-efficient reconfigurable I/Os.
Big memory sea (nonvolatile memory)
Thousands small Accelerator cores
High-bandwidth, energy-efficient reconfigurable I/Os
Figure 1.6 Ideal logic-in-memory architecture
12
ReRAM-based machine learning
Binary address
Wordline decoder
Considering the leakage reduction at the same time, logic-in-memory architectures that are associated with NVM are presented in [43–47]. Figure 1.7 shows a logic-inmemory architecture by building CMOS circuits with storage devices. The example shows a full-adder with both sum and carry logic. However, such a cell-level inmemory architecture has two major problems. First, the CMOS-based logic is built in memory so that it is difficult to be reconfigured. Second, it can only perform lowcomplex logic operations. In the example, it needs 22 transistors for a simple 1-bit adder. The complexity of memory will be thus greatly increased. Another in-memory architecture to investigate at the block level in distributed fashion is illustrated in Figure 1.8, which is more effective for traffic reduction. A memory data is usually organized in H-tree fashion, and the data block can be the data array or a number of data arrays that belong to same “H-tree” branch. Instead
Bitline decode
Column multiplexer Sense amplifiers
Output WL
WL′
WL
WL′
VDD CLK
CLK
BL
B Cin
B′ C′in A
A′
CLK
CLK
B′ C′in
B Cin
B′ Cin
A
B C′in
A′
BL′
GND Carry logic
Sum logic
Figure 1.7 IMC architecture at memory-cell level
In-memory logic
Nonvolatile storage
Introduction
13
External processor Memory Data/address/command I/O
Local data/logic pair
Data array
In-memory logic
Local data path for in-memory logic
Figure 1.8 IMC architecture at memory block level
of inserting an in-memory logic at memory-cell level inside the data array, the architecture in Figure 1.8 pairs each block of data with in-memory logic (accelerators). Different from the cell level in-memory architecture, the accelerators can be made with higher complexity, and the number of accelerators for each data block can also be customized. The data flow of the block-level in-memory architecture is to read out data from data block to in-memory logic, which performs particular functionality and then writes back the result. The data also needs to be stored in assigned blocks, but it is much more flexible than that of cell level in-memory architecture. The block-level in-memory architecture is very effective to reduce communication traffic between memory and processor. This is because significant operands reduction can be achieved by deploying accelerator with high-level functionality. For example, for face recognition in image processing application domain, instead of transmitting a whole image to obtain a Boolean result, the result can be directly gained through inmemory logic. In other words, the block-level in-memory architecture is suitable for big-data-driven applications where traffic reduction is more beneficial than latency reduction. In this architecture, the nonvolatile ReRAM is intensively used. Both the memory block and logic block in each pair are purely implemented by ReRAM devices. In addition, energy-efficient in-memory logic units are deployed in the external processor to execute instructions that cannot be accelerated by in-memory logic.
14
ReRAM-based machine learning
1.2 Challenges and contributions Future cyber-physical systems require efficient real-time data analytics [48–52] with applications in robotics, brain–computer interface as well as autonomous vehicles. The recent works in [53,54] have shown a great potential for ML with significant reduced training time for real-time data analytics. Hardware-based accelerator is currently practiced to assist ML applications. In traditional hardware accelerator, there is intensive data migration between memory and logic [55,56] causing both bandwidth wall and powerwall. Therefore, for dataoriented computation, it is beneficial to place logic accelerators as close as possible to the memory to alleviate the I/O communication overhead [57]. The cell-level IMC is proposed in [43], where simple logic circuits are embedded among memory arrays. Nevertheless, the according in-memory logic that is equipped in memory cannot be made for complex logic function, and also the utilization efficiency is low as logic cannot be shared among memory cells. In addition, there is significant memory leakage power in CMOS-based technology. Emerging ReRAM [58–64] has shown great potential to be the solution for dataintensive applications. Besides the minimized leakage power due to nonvolatility, ReRAM in crossbar structure has been exploited as computational elements [59,65]. As such, both memory and logic components can be realized in a power- and areaefficient manner. More importantly, it can provide a true in-memory logic-memory integration architecture without using I/Os. Nevertheless, the previous ReRAMcrossbar based computation is mainly based on an analog fashion with multilevel values [66] or spike timing-dependent plasticity (STDP) [67]. Though it improves the computational capacity, the serious nonuniformity of ReRAM-crossbar at nanoscale limits its wide applications for accurate and repeated data analytics. Moreover, there is significant power consumption from additional AD-conversion and I/Os mentioned in [67]. Besides ReRAM-crossbar, coupled ReRAM-oscillator can be utilized for analog computing. By mapping fixed-point values to voltages, coupled ReRAMoscillator can perform an L2-norm calculation. Significant speed-up and energy efficiency can be achieved due to the direct analog computing. Although the nonvolatile IMC seems to be a promising solution for future dataoriented storage and computing system, the recent emerging ReRAM are still in their infancy due to the following unresolved challenges:
●
●
From circuit level, although there exist some simulators to perform hybrid NVMCMOS circuit simulation [63,64], the ReRAM-based computation operations have not been fully developed yet. Without the detailed IMC implementation on ReRAM, it is hard to perform a real application on such an ReRAM-based nonvolatile in-memory accelerator. From architecture level, it is necessary to explore a detailed implementation for the in-memory architecture. Controllers for conventional memory are not optimal for the ReRAM-based IMC system. The CMOS-based control bus and its communication protocol need to be redefined for such a novel architecture.
Introduction ●
15
From system level, it is necessary to find out what ML algorithms can be mapped on the in-memory accelerator. For different operations or algorithms, we need to design different accelerators in order to achieve a higher parallelism and a better energy efficiency.
This book presents the idea of exploring the development of ReRAM-based logic operations and NVM in-memory architecture, as well as mapping details of various ML algorithms on such an architecture. For the ReRAM-based IMC implementation, the target is to develop detailed mappings for different operations on ReRAM. We proposed two kinds of ReRAM-based structures for operation accelerating: ReRAMcrossbar and ReRAM oscillator network. Particularly, this book has summarized the following advancements: ●
●
●
First, the NVM-SPICE is used for circuit-level simulation. The model of drifttype ReRAM is introduced with simulation results. In addition, the measurement results of forming, SET and RESET processes are also shown. For the diffusivetype ReRAM, a sub-circuit model is used for simulation. Second, we introduce a binary ReRAM-crossbar for matrix-vector multiplication. Due to the nonuniformity of ReRAM device, the proposed binary ReRAM-crossbar can perform the computation more accurate compared to traditional ReRAM-crossbar in analogue fashion. In addition, three kinds of crossbar structures including passive array, one-transistor-one ReRAM (1T1R) and one-selector-one ReRAM (1S1R) are discussed. Third, we describe an ReRAM-based coupled oscillator network for L2-norm calculation. A simple oscillator can be built based on the ON-OFF switching of diffusive ReRAM or the forming process of drift ReRAM. When the basic oscillators form a network, it can perform an L2-norm calculation based on the fitted simulation results.
For the in-memory accelerator, the target is to determine the design of architecture. For the architecture with only ReRAM as computing engine, we design the auxiliary CMOS logic and define the communication protocol between the processor and memory in detail. For the architecture with both ReRAM and CMOS as computing engine, we develop 3D CMOS-ReRAM architecture so that a higher throughput and parallelism can be achieved by introducing the following two topics: ●
●
First, a distributed in-memory computing architecture (XIMA) is described. Despite traditional store and load, we define two more commands: start and wait. In addition, all the logic blocks and data blocks are formed in pairs so that the delay of data-processing communication can be minimized. CMOS logic such as instruction queue and decoders are also designed. Second, design space exploration of 3D CMOS-ReRAM architectures is presented. For single-layer 3D architecture, ReRAM devices are via connecting top-layer WLs and bottom-layer BLs. All the other CMOS logics are implemented in the bottom layer. For multilayer 3D architecture, we use two ReRAM layers as data buffer and logic implementation, respectively. In addition, TSVs are used to connect different layers.
16
ReRAM-based machine learning
For the data analytics mapping on ReRAM-based architecture, the target is to focus on detailed accelerator design for specific applications. In addition, the algorithm requires to be optimized so that all the operations can be fully mapped on the accelerator. We present case studies with several ML algorithms on distributed inmemory architecture (XIMA) and 3D CMOS-ReRAM architectures with significant improvement on bandwidth and energy efficiency. ●
●
First, for the XIMA, we have accelerated three ML algorithms. The learning and inference procedures of single-layer feedforward neural network (SLFN) have been optimized and partially mapped on the passive binary ReRAM-crossbar. In addition, we mapped the binary convolutional neural network (BCNN) on both passive array and 1S1R array with different mapping schemes. Moreover, L2norm gradient-based learning and inference are also implemented on an ReRAM network with both crossbar and coupled oscillators. Second, for 3D CMOS-ReRAM architecture, we also mapped the optimized SLFN algorithm. All the operations in learning and inference stages are implemented so that it can achieve the online learning. In addition, tensorized neural network (TNN) is mapped on both single-layer and multilayer accelerators with different mapping scheme.
For every ML algorithm mapped on XIMA and 3D CMOS-ReRAM accelerator, we evaluate their performance in device, architecture and system levels. All the implementations show higher throughput, bandwidth and parallelism with better energy efficiency.
1.3 Book organization This book covers the entire flow from device, architecture and learning algorithm to system perspectives for emerging ReRAM. It is organized into the following chapters. The first chapter introduces the need and motivation to deploy ML algorithms for data-intensive computations. It also introduces the challenges and need for hardware for ML processing. In the second chapter, we present emerging memory devices that can support IMC and need for IMC are introduced in this chapter. Further, a comprehensive comparison and analysis of different IMC architectures are presented. The architectures include using traditional DRAM to emerging devices such as ReRAM and memristors. In the third chapter, modeling of ReRAM devices is presented followed by design of crossbar structures that are primarily used for matrix multiplication in ML applications. In addition, design of multibit storage for ReRAMs and other logic functional units with ReRAMs is presented. The fourth chapter provides the basic information on ML algorithms, computational complexity and the types of operations involved in the ML algorithms, including least-square method (support vector machine (SVM), extreme learning machine (ELM)) and deep learning (CNN, TNN). Trained quantization (binary, ternary, eightbit, etc.) will be addressed in this chapter. Additionally, the training mechanisms for
Introduction
17
deep CNNs and TNNs are focused in this chapter, as they are the time-consuming operations and the performance depends on the training efficiency. The fifth chapter introduces multiple strategies to map the ML designs on to hardware accelerators. Designing of ReRAM crossbar-based IMC architectures to accelerate ML applications is presented. Further, to accommodate deep neural networks (DNNs) and CNNs, a 3D stacking-based CMOS-ReRAM architecture is introduced. The main focus of this chapter is on the architectural aspects and circuit designs to support the data-intensive computations with high speed. In Chapter 6, we introduce techniques to map the ML applications for ReRAM and the major challenges and the scope for improvement. The mapping techniques for different ML algorithms for different architectures such as ReRAM cross with traditional 2D and 3D architectures are illustrated. Case studies for ReRAM-based ML acceleration for different scenarios are presented. Chapter 7 presents ResNet under different scenarios such as DNN with quantization, ReRAM-based acceleration for ResNet. In addition to the ResNet, this book presents ReRAM-based accelerator for compressive sensing application in Chapter 8. For this part, different scenarios such as iterative heuristic algorithm and row generation algorithms are presented. All the case study chapters are presented with substantial amount of experimental results and analysis that aid in replicating the experiments and also prove the efficiency of ReRAM accelerators and its prominence in the design of future hardware accelerators. The last chapter summarizes the ReRAM accelerators and its impact on the ML applications. Furthermore, future directions with ReRAM devices to further enhance the ML computations are presented.
This page intentionally left blank
Chapter 2
The need of in-memory computing
2.1 Introduction In 1974, Dennard predicted the reduction of switching power with the downward scaling of feature size of transistor [68], and this trend ended around 2004. The operating voltage and current reduction of transistor stagnated despite the ever-shrinking size of transistor because of the increasing leakage current. The increase of the operating frequency and the transistor density further exacerbates the power consumption of a computing system, which generates a large amount of heat and impedes the power efficiency. Another challenge for the power consumption is the intensive data transfer between data processing and memory units in conventional Von Neumann or Princeton computing architecture, so called the “memory wall.” In this computing paradigm, the computing unit can only process one task at a certain interval and wait for memory to update its results, because both data and instructions are stored in the same memory space, which greatly limits the throughput and causes idle power consumption. Although mechanisms like cache and branch prediction can partially eliminate the issues, the “memory wall” still poises a grand challenge for the massive data interchanging in modern processor technology. To break “memory wall,” in-memory processing has been studied since 2000s and regarded as a promising way to reduce redundant data movement between memory and processing unit and decrease power consumption. The concept has been implemented with different hardware tools, e.g., 3D-stack dynamic random access memory (DRAM) [69] and embedded FLASH [70]. Software solutions like pruning, quantization and mixed precision topologies are implemented to reduce the intensity of signal interchanging. The biological neural network (NN) outperforms conventional Von Neumann computing architecture based on central processing unit (CPU) or graphics processing unit (GPU) in term of power efficiency, which consumes only 1–100 fJ per synaptic event when performing intense parallel computing [71,72]. Currently, the synaptic resistive random access memory (ReRAM) has been successfully demonstrated to exhibit low power consumption ( Vr
(3.8)
where Vth is the threshold voltage of the ReRAM. Because of the high density of ReRAM device, one can build a crossbar structure as the array of ReRAM [59,140–144]. Such crossbar structure can be utilized as memory for high-density data storage. The memory array can be read or written by controlling the voltage of wordlines (WLs) and bitlines (BLs). For example, we can apply Vw /2 on the ith WL and −Vw /2 on the jth BL to write data into the ReRAM cell on ith row, jth column.
The background of ReRAM devices
55
3.2.1 Analog and digitized ReRAM crossbar This section introduces the traditional analog ReRAM crossbar and the digitized ReRAM crossbar in the following. The I/O interface and the number of ReRAM device resistance value are different in these two structures.
3.2.1.1 Traditional analog ReRAM crossbar ReRAM is an emerging nonvolatile memory based on two-terminal junction devices, whose resistance can be controlled by the integral of externally applied currents. The fabric of crossbar intrinsically supports matrix–vector multiplication where vector is represented by row input voltage levels and matrix is denoted by the mesh of ReRAM resistances. In order to calculate the matrix–vector multiplication y = x. As shown in Figure 3.11, by configuring into the ReRAM crossbar and setting x as the crossbar input, the analog computation can be achieved by ReRAM crossbar directly with output y. However, such an analog ReRAM-crossbar has two major drawbacks. First, the programming of continuous-valued ReRAM resistance is practically challenging due to large ReRAM process variation. Specifically, the ReRAM resistance is determined by the integral of current flowing through the device, which leads to a switching curve as shown in Figure 3.12(a). With the process variation, the curve may shift and leave intermediate values very unreliable to program, as shown in Figure 3.12(b). Second, the analog-to-digital converter (ADC) and digital-to-analog converter (DAC) are both time-consuming and power-consuming. In our simulation, the analog-to-digital and digital-to-analog conversion may consume up to 85.5% of total operation energy in 65 nm as shown in Figure 3.13.
3.2.1.2 Digitalized ReRAM crossbar To overcome the aforementioned issues, a fully digitalized ReRAM crossbar for matrix-vector multiplication. First, as ON-state and OFF-state are much more reliable
x1
DAC
x2
DAC
xn
G11
G12
G1m
G21
G22
G2m
V2
Gn1
Gn2
Gnm
Vm
V1
ADC
y1
ADC
y2
ADC
ym
DAC
Transimpedance amplifiers (TIA)
Figure 3.11 Traditional analog-fashion ReRAM crossbar with ADC and DAC
56
ReRAM-based machine learning Intermediate states Inaccuracy
Resistance
t(Roff/2) Off-state Inaccurate resistance programming
On-state
On-state
Off-state Target resistance
Time (a)
(b)
Figure 3.12 (a) Switching curve of ReRAM under device variations. (b) Programming inaccuracy for different ReRAM target resistances
0.42%
14.51% 4.88%
31.45% ADC DAC RRAM
ADC DAC RRAM 68.13%
80.61% (a)
(b)
Figure 3.13 (a) Power consumption of analog-fashion ReRAM crossbar. (b) Area consumption of analog-fashion ReRAM crossbar
than intermediate values shown in Figure 3.12, only binary values of ReRAM are allowed to reduce the inaccuracy of ReRAM programming. Second, a pure digital interface without A/D conversion is deployed. j i In ReRAM crossbar, We use Vwl and Vbl to denote voltage on ith WL and jth BL. RLRS and RHRS denote the resistance of LRS and HRS, respectively. In each sense amplifier (SA), there is a sense resistor Rs with fixed and small resistance. The relation among these three resistance is RHRS RLRS Rs . Thus, the voltage on jth BL can be presented by j
Vbl =
m
gij Vwli Rs
i=1
where gij is the conductance of Rij .
(3.9)
The background of ReRAM devices
57
The key idea behind digitalized crossbar is the use of comparators. As each column output voltage for analog crossbar is continuous-valued, comparators are used to digitize it according to the reference threshold applied to SA j j 1, if Vbl ≥ Vth Oj = (3.10) j j 0, if Vbl < Vth However, the issue that rises due to the digitalization of analog voltage value is the loss of information. To overcome this, three techniques can be applied. First, multi-thresholds are used to increase the quantization level so that more information can be preserved. Second, the multiplication operation is decomposed into three suboperations that binary crossbar can well tackle. Third, the thresholds are delicately selected at the region that most information can be preserved after the digitalization. Compared to the traditional analog ReRAM-crossbar [145,146], the advantages of the digital ReRAM-crossbar are the following: ●
●
●
It has better programming accuracy of the ReRAM device under process variation with no additional ADC conversation as well. Since only LRS and HRS are configured in digital ReRAM crossbar, it does not require high HRS/LRS ratio so that low-power (high LRS) device [139] can be applied. In addition, the wire resistance affects little on IR-drop when using the low-power ReRAM device. The binary input voltage has a better robustness on the IR-drop in large-size crossbar.
However, the area and energy consumption of ReRAM devices will increase in the digital ReRAM-crossbar implementation. It will also be discussed in Section 5.
3.2.2 Connection of ReRAM crossbar There are three approaches for the connection of ReRAM crossbar: direct-connected ReRAM, one-transistor-one-ReRAM (1T1R), and 1S1R. These approaches are described in-detail along with the differences among them.
3.2.2.1 Direct-connected ReRAM Direct-connected ReRAM, also known as the passive ReRAM array, which is a crossbar structure with only ReRAM devices. Every WL and BL are connected with a single drift-type ReRAM. The structure of direct-connected ReRAM is shown in Figure 3.14(a), and here we use ith WL (WLi ) and jth BL (BLj ) as an example. To perform a write operation, one needs to apply a write voltage Vwrite between WLi and BLj . In this case, the ReRAM connecting them (Ri,j ) can be precisely configured. However, a problem exists in this structure that sneak path may occur. In this example, the purple path is the ideal current path we want to have, but there may be another path identified by green. This path consists of Ri−1,j , Ri−1,j−1 and Ri,j−1 so that we can see a Vwrite is also applied on the serial of these three ReRAMs. In some of the cases, two of them are in LRS and one of them is in HRS, causing a very large voltage applied on the ReRAM device in HRS. In this case, the status of this ReRAM may also be configured, which we do not want it happens.
ReRAM-based machine learning
…
Ri,j
… VG,1
…
…
Output
Output
(a)
Output
Output
Output
Output
(b)
Analog Multibit interface input 1-bit input
Output
Output
(c) Analog interface
DAC
…
…
…
…
VG,n
BLj
Digital interface
DiffusiveReRAM
…
…
Output
…
…
…
…
…
VG,2
…
…
Gate control
Transistor Wordline input voltages
WLi
…
Wordline input voltages
DriftReRAM
Sneak path
Target path
Wordline input voltages
58
Digital interface Vth
ADC
Multibit output 1-bit Output
Output module module Output
Input module (d)
Figure 3.14 Comparison of ReRAM-crossbar implementation among (a) direct connection with sneak path; (b) 1T1R with extra control signals; (c) proposed sneak-path-free 1S1R without extra control signals; (d) comparison of analog and digital interfaces
To overcome the sneak-path issue, a solution is proposed in [147–149] that we can applied a constant voltage on the rest of WLs and BLs. Here, the target WL and BL are applied with 2/Vwrite and −2/Vwrite , respectively. Meanwhile, we also apply 0 V on all the inactive lines. This solution has two requirements for the device and controller design: ●
●
The nonlinearity of ReRAM device should be very good since the device cannot be configured with a 2/Vwrite drop. The external CMOS controller needs to output multiple voltages for all lines at the same time. In addition, the amplitude of voltage of all the lines should be independent.
3.2.2.2 One-transistor-one-ReRAM In 1T1R structure, an NMOS transistor in connected with each ReRAM device serially indicated in [145]. As a result, each WL is connected to a BL with one transistor and one ReRAM, as shown in Figure 3.14(b). The NMOS transistor is used as a voltagecontrolled selector so that we can configure the connectivity of each ReRAM cell. The major benefit of 1T1R is to avoid the sneak path in the direct-connected structure.
The background of ReRAM devices
59
In Figure 3.14(b), we use one signal to control the transistors connecting to the same WL. It is clear that we only need to apply voltage on the target WLi , BLj and only put VG,i to high, only the target cell can be configured. The metric of this structure can be summarized as follows: ●
●
Low requirement on the nonlinearity of ReRAM device because voltage is only applied on the target cell. The controller design is much more easier than the direct-connected structure.
However, the 1T1R structure still has a disadvantage that the crossbar array area depends on the transistor, which is not scalable compared to ReRAM device. Therefore, it is not a good option when the size of ReRAM crossbar is large.
3.2.2.3 One-selector-one-ReRAM A sneak-path-free digital ReRAM-crossbar is shown in Figure 3.14(c). In this architecture, every drift-ReRAM is connected in serial with a diffusive ReRAM as selector. Compared to the previous two implementations, the 1S1R can remain good scalability and eliminate sneak path without additional controllers. Compared to [150], the Ron of selector can only affect little to the operation on digital ReRAM. Since Roff RHRS as shown in Figure 3.10, the static resistance of 1S1R is almost equal to Roff . In addition, the majority of applied voltage is on the selector device when it is OFF. Since each sneak path will connect to at least three 1S1Rs, the operating voltage V only needs to satisfy the following equation to eliminate the sneak path: 3Vth,on > V > Vth,on
(3.11)
We will further discuss about the simulation results regarding the sneak-path-free effect in Section 7.5.
3.3 ReRAM-based oscillator Oscillatory behavior in ReRAM device has been observed in [151,152], and it is performed by the forming process of a drift-type ReRAM. The recent diffusive-type ReRAM [137,138] shows the feasibility as oscillators. The ReRAM-based oscillator can be built as shown in Figure 3.15(a). In this ReRAM-based oscillator, a physical resistor Rs is in series with the ReRAM device. The output voltage will depend on the proportion of Rs and ReRAM resistance RM . With the charge/discharge of capacitor ·RM CM , the output voltage Vout will be approaching VRDD . To make the oscillation work, M +Rs we have the following requirements for the ReRAM device in the structure: ●
●
At a relatively high voltage Vh , the resistance of the ReRAM RM decreases significantly to Rl . If we drop the voltage to a low value Vl , RM will increase to Rh .
60
ReRAM-based machine learning 800 700
Rs
600
V (mV)
VDD
Vout
500 400 300 200
RM
CM
100 0 0
250
500
750
1,000
Time (ns) (a)
(b) 250 6.5
Frequency (MHZ)
Mag(E-3)
208 167 125 83.3 41.7 0
6.4 6.3 6.2 6.1 6.0 5.9
5
6.5
8.0
90
Frequency (MHZ) (c)
95
100
105
110
Resistance (kΩ) (d)
Figure 3.15 (a) ReRAM-based oscillator schematic; (b) simulation result of oscillator; (c) DFT analysis of oscillator; (d) oscillatory frequencies vs. different Rs values
Therefore, we can build up the oscillation if all the parameters in Figure 3.15(a) are well-fitted, following Rl Rh < Vl < Vh < Rl + R s Rh + R s
(3.12)
where Rl and Rh are, respectively, the low resistance and high resistance of the ReRAM device. As a result, RM and Vout cannot be in a static state with the condition stated in (3.12). When the ReRAM turns from Rl to Rh or vice versa, a negative resistance effect is performed. Moreover, the oscillatory frequency is determined by the value of Rs . The ReRAM model for oscillator according to [153] is built for case study, and one can observe the oscillation result in Figure 3.15(b). Figure 3.15(c) shows the Discrete Fourier transform (DFT) analysis of Figure 3.15(b). Figure 3.15(d) shows the oscillatory frequencies vs. different Rs values. With higher Rs , there is less current passing through the oscillator such that the charge/discharge will be slower, resulting in lower oscillatory frequency.
The background of ReRAM devices
61
3.4 Write-in scheme for multibit ReRAM storage 3.4.1 ReRAM data storage Electrical engineers are quite familiar with circuits which include the three basic passive circuit elements—resistor, capacitor and inductor. However, for the first time, in 1971 Leon Chua proposed a fourth circuit element, describing the relation between charge (q) and flux (φ) [60]. The resistance of the so-called memristor device depends on the total charge passed through the device [60,154]. In this chapter, we use the terms ReRAM and memristor interchangeably. The device characteristics of the memristor open the floor for many new fields of application [155–159]. One of the most important applications is, as expected, using it as a memory element. For example, [160–162] and others have invested considerable effort in studying the design of memristor-based memory units. The main advantage of using memristors is given by the fact that in a modern chip, the number of transistors required to store data (e.g., in an SRAM) has a significant—and increasing—impact on the total transistor count [163]. In consequence, implementation of new memory architectures by using memristors would decrease the total amount of device leakage current dramatically [161]. Since HP developed the first passive memristor in 2008 [62], there have been works on single-bit [160–162,164] and multibit [165,166] memory storages. The unique φ–q characteristic of memeristors, however, leads to a nonlinear v–i characteristic, which makes it difficult to determine the pulse size for achieving a certain state [165,166]. Therefore, it has been seen as a problem for multilevel memory designs [165,166]. To overcome this problem, researchers have developed methods, some fairly complicated, which use analog circuits and Opamps for both reading and writing [165,166]. Using Opamps and analog circuits for readout seems to be inevitable [160,165,166]. However, as we will present, the aforementioned characteristic can be taken advantage of, to store more than one bit in memristors, using circuits which are compatible with digital designs. These circuits are hence less complex. Although digital bit streams have been used for setting synaptic values [167], the exact stored value and recovery of the stored value are not of high importance in such applications. In memory applications, on the other hand, this is of paramount importance. Therefore, in this chapter, we discuss the reliability of the storage and readout, as well as how encoding can help improving these parameters. The rest of this chapter is organized as the following: In the next section, we briefly review the aforementioned unique characteristics of memristor and show how it can be taken advantage of, in order to store more than one bit in each memristor. In Section 3.4.3, we present the digital writing method for storing two bits of data on a single memristor, as well as the readout circuit. Then, in Section 3.4.4, we show the results of our simulations which confirm the feasibility of the proposed approach. In Section 3.4.5, we improve the reliability of the storage, and the number of stored bits, using different proposed encoding schemes, especially regarding practical implementations and further verification of the proposed approach.
62
ReRAM-based machine learning
3.4.2 Multi-threshold resistance for data storage It is a well-established phenomenon in the literature that the resistance of a memristor depends on the charge flown through it [60,168]. This can be modeled in various ways, one of the most prominent of which is [168]: Ron − Roff (3.13) e−4km (q(t)+q0 ) + 1 where Roff and Ron are the maximum and minimum resistance of the memristor, and q(t) is the charge flown through the memristor with the initial value of q0 . Finally, km is a constant which represents physical characteristics of the device such as doping of the semiconductor and its size. According to (3.13), not only the resistance, but also changes of the resistance in a memristor due to identical voltage pulses depend on its current state [165,169]. This characteristic seen as a hardship [165,166] can be turned into an opportunity based on the following theorem. Theorem: Distinct output resistances correspond to distinct input patterns, which allows storage of more than one bit in a single memristor. Short Proof∗ : Assume two distinct states A and B, seen in Figure 3.16, with unequal resistance of Rx , where x ⊂ {A, B}. Now, we are interested in finding out the ratio of resistance change at these two points, due to the application of an input voltage pulse with the amplitude of V and width of T . Based on Ohm’s law, we have R(q(t)) = Roff +
dq (3.14) dt where R(q(t)) of the memristor is given by (3.13). Solving this equation is rather complicated and out of the scope of this chapter. Therefore, in a simplified manner, we estimate the changes of resistance at each point by its slope, mx . In other words, in the vicinity of point x, we estimate (3.13) by V = R(q(t))
R(q) = −mx q + Rx
(3.15)
Roff R(q) A0 RA
A
mA A01 A10
RB Ron
A1 –q
0
mB
B +q
Figure 3.16 Resistance of a memristor as a function of the charge stored in it ∗
The full proof is provided in [169]. We kindly ask the readers to refer to that paper for more information on details.
The background of ReRAM devices
63
By substituting (3.15) into (3.14), and integrating over the period, T , we have [169] → Vdt = Rdq = (−mx q + Rx )dq
T
dq → 0 Vdt = 0 x (−mx q + Rx )dq →
−mx (dqx )2 2
VT =
+ Rx dqx
(3.16)
Solving (3.16), we obtain dqx =
Rx Kx mx
where
(3.17)
Kx = 1 +
1−
2mx VT Rx 2
(3.18)
x On the other hand, we can easily infer from (3.15) that dR = −mx . Based on this, the dqx ratio of resistance change due to charge change at point A and B is
dRA dRB mA dRA RA KA / = → = dqA dqB mB dRB RB KB
(3.19)
This shows how the change of resistance of memristors at different stages due to identical pulses depends on its state (resistance) at the time. This ratio could be further simplified and approximated to the resistance ratio only [169]. Now, assuming that a positive pulse presents “1” and a negative pulse presents “0,” if a memristor is fed by a positive input pulse followed by a negative input pulse (“10”), it will reach a different state (“A10 ”), compared to the case where the negative pulse is applied before the positive pulse (“01” and hence A01 ). The rationale behind it is that the resistance change due to an initial negative pulse (A to A0 ) is different compared to the resistance change due to the same pulse, once the state of memristor has changed due to a previous positive pulse (A1 to A10 ). The same holds for changes due to positive pulses in two different states (A0 to A01 versus A to A1 ). Therefore, we can conclude that, since the states of the memristor after application of “01” and “10” are different, based on the state of the memristor, its original input can be retrieved [169]. This will be verified in the next section.
3.4.3 Write and read Based on the aforementioned concept, in this section, we present our read and write method and simulate them in Section 3.4.4, to test the feasibility of successfully recovering a data value which was stored through digital streaming. For the array architecture, isolation and access to the memory cell, a 1T1R architecture similar to [166] is assumed. Since the Opamp (comparator) is the crucial part of the circuit determining the feasibility and reliability of this implementation, we have used a model of an existing off-the-shelf Opamp, namely LT1012, which takes into account nonideal characteristics such as offset.
64
ReRAM-based machine learning
3.4.3.1 Write-in method As previously discussed in order to distinguish the “1” input and “0” input, respectively, positive and negative pulses are used, which can be implemented through simple switches to reference voltages. This eliminates the need for the complex circuit proposed in [165] or the DAC used in [166]. The magnitude of these pulses for the current experiment is 0.5 V and the pulse widths are 10 ms. The inputs were applied based on the most-significant-bit first and the least-significant-bit last. Table 3.2 shows the data to be stored and the resulting state of the memristor (resistance) due to these inputs. This table shows that, as expected, different input data patterns lead to different states for the memristor. Similar to [166], the thermometer readout codes at this column can be later converted to the normal binary representation (similar to the left-most column) using simple digital logic circuits. Only one instance of such a circuit will be necessary for each memory block containing hundreds to thousands of memristor cells. Therefore, the cost of the additional hardware is negligible.
3.4.3.2 Readout method For retrieving the stored data, [160] uses an Opamp-based circuit, [166] uses an ADC, and [165] uses a circuit inspired by ADCs. In this chapter, as shown in Figure 3.17, we also use a circuit inspired by flash ADCs. In this circuit, after the inputs are applied and the data are stored in the memristor, the input source is disconnected. Then, after applying a controlled current pulse (here a rectangular pulse with the magnitude of 1 μA and width of 1 ms), the resulting voltage across the memristor is buffered and fed to three comparator circuits (only one of which is shown at the bottom of Figure 3.17, for simplicity). The voltage across the memristor is then compared with the voltage produced by applying a similar current pulse to predetermined resistance values. If the voltage across the predetermined resistors (Ri ) is greater than the voltage read from the memristor, the comparator will flip to high output and a “1” will be stored in the respective flip-flop output (RDi ). This implies that the resistance of the cell is below the respective resistor, Ri . We note that the readout current pulse is chosen to be significantly smaller than input pulses, so that it would have a smaller effect on the state of the memristor. Nevertheless, however small this effect is, it needs to be compensated for. Therefore, similar to [160,166], by applying another controlled pulse in the reverse direction, this Table 3.2 Data to be stored on the memristor, resulting states and expected unencoded readouts Data
Rmem (k)
Exp. readout
“00” “01” “10” “11”
5.213 5.108 4.890 4.778
“000” “001” “011” “111”
The background of ReRAM devices
Vin
Vdd lread
65
Vbuff
Vss Vdd
lread
Ri
D
Vbuff Vss
Clk
SET Q
RDi
CLR Q
Figure 3.17 For reading out, Vbuff is fed to three comparators connected to D flip-flops which provide RDi bits (for simplicity, only one set is depicted)
effect can be compensated for. The advantage of using current pulses, compared to the voltage pulses used in [160,166], is that the effect of positive and negative current pulses on the resistance change will be similar, whereas as proved in Section 3.4.2, it is not the case for voltage pulses. That is, a negative voltage pulse, due to inherent characteristics of the memristor and its nonlinear voltage–current relationship, cannot precisely compensate a positive voltage pulse and will leave residual extra charges (positive or negative). Therefore, the state of memristor needs to be checked and more often refreshed to compensate the residual charges as well. Using current sources, as proposed here, the complex method of readout effect elimination is reduced to an automatic compensation through a reverse current pulse, which is considerably simpler than the one used in [165]. Automatic refreshing may lead to extra power consumption; however, it eliminates the state-read operation or counting process for the refreshing proposed in [166]. Therefore, it is going to be simpler. We note that this technique may not prove efficient for all types of memristor technologies (e.g., for non-TiO2 memristors) or scenarios (e.g., for power-constrained applications). Therefore, in order to compensate for the destructive effect of uncompensated readouts (for power critical applications or where this technique is not efficient), other scenarios, as discussed in [166], can be employed to compensate for the readout effect.
3.4.4 Validation In order to evaluate the practicality of the proposed approach in storing data in memristors, we have simulated the proposed circuit in LTSPICE. We have used the most prominent model (namely, Biolek’s [168]) of TiO2 memristors, described by
66
ReRAM-based machine learning 1
RD1 Vin Voltage (mV)
0.6
Voltage (V)
0.5
0
0.4
RD1 Vin
0.2 0 3.07 3.08 3.09
3.1
3.11 3.12 3.13
Time (s)
–0.5
–1 0
0.5
1
1.5
Time (s)
2
2.5
3
3.5
Figure 3.18 Simulation results for “01” input. Shortly after applying the readout current at 3.1 s (zoomed in), the output of RD1 flips to “1” representing the stored data. RD0 and RD2 remained zero and are not shown for simplicity
(3.13). The following values were used for the model parameters (as given by the authors [168]); Ron = 100 , Roff = 10 k, km = 10, 000 and Rinit = 5 k. Input signals last 10 and 1 ms, respectively, for input voltages and readout current pulses. The amplitudes are ±0.5 V for “1” and “0” input pulses and −1 μA for readout current pulses. The value of Ri resistors used in this set of simulations are 5,170, 5,130 and 5,010 for i =1, 2 and 3, respectively. Comparators and their precision play a crucial role in the practical implementation of the readout circuit, which determine to what level the small differences between states can be distinguished. To account for their nonidealities, we have used the model for an off-the-shelf Opamp, namely LT1012 (with 25 μV maximum offset). This ascertains that the assumed values for parameters such as offset are realistic and accounted, i.e., this circuit can—in practice also—perform as expected. Figure 3.18 shows the result of simulation for a sample case, namely “01” input. In this figure, the blue curve shows the voltage across the memristor (namely Vin ) and the black curve shows the RD1 output. For simplicity, RD2 and RD3 which as expected remained zero during the whole simulation are not shown. As we can see, shortly after applying the readout current at 3.1 s (zoomed-in area) the output of RD1 turns to “1.” This and other simulations for all outputs satisfied the expected outputs as shown in Table 3.2, where the expected outputs represent “RD3 & RD2 & RD1 ” combination. Once these values are obtained, different inputs are distinguished properly and with simple logic circuits, these outputs can be turned into binary values once again. This proof of concept confirms the feasibility of digital streaming approach for storing, and successfully retrieving two bits of information on a single memristor.
The background of ReRAM devices
67
Table 3.3 Resistance for memristors with different ranges of memristance. M1 : 0.1–10 k and M2 : 0.5–100 k Input
RM1 (k)
RM2 (k)
“000” “001” “010” “011” “100” “101” “110” “111”
7.2453 5.8991 5.8997 3.9089 5.8990 3.9078 3.9088 1.0396
47.2428 28.9629 28.9710 5.0033 28.9600 4.9898 5.0014 0.5000
3.4.5 Encoding and 3-bit storage In this section, we present the impact of the range of the memristor and the encoding of data on the reliability of the readout.
3.4.5.1 Exploration of the memristance range Trying to store 3 bits or more in a single memristor can have a negative impact on the reliability of the readout, due to smaller differences between resistances at different states. Expanding the range of the memristor (between Ron and Roff ) helps in reducing the relative sensitivity. Although the relative difference in memristance between inputs with the same number of ones or zeros can be thus improved, this improvement is not necessarily proportional to the expansion of the range. For example, in a memristor with a range of 100 –10 k, if the difference between “10” and “01” is 10 , in a memristor with a range of 1–100 k, the difference—according to our simulations—can be 17 . This is validated in our simulations, where two memristors are considered: memristor 1 (M1 ) with the range of 0.1–10 k and memristor 2 (M2 ) with the range of 0.5–100 k. The pulse width is set to 100 ms, and the amplitudes are set to ±0.5 V for “1” and “0” inputs. Initial state of memristor is assumed to be 5 k. Table 3.3 presents the memristance value for different 3-bit data patterns. For instance, the memristance difference between “011” and “101” is 1 for M1 , whereas it is 14 for M2 . This indicates that inputs with the same number of ones or zeros can be differentiated easier in the memristor with a wider resistance range. Although in this way the difference is increased, it is yet small and requires complex complementary circuits, such as the one used in [165], to set the values precisely. It also needs precise comparators, such as the one assumed in [166], to read the values out. To alleviate the problem of close resistance at different states, and in order to improve reliability without using analog write-in circuits or extra circuits with extensive complexity for the readout operation, we propose two encoding
68
ReRAM-based machine learning
schemes. These techniques lead to a significantly higher relative memristance difference between inputs with the same number of ones and zeros. As the number of memristors are not increased with encoding, the area of the memory remains constant, except for the addition of an encoder for the whole system (which is a memory block with hundreds to thousands of memory cells). Therefore, the area used for the encoder is considerably negligible.
3.4.5.2 Uniform input encoding In the first scheme, we employ a simple encoding method. In this technique, we append the most significant bit (MSB) to the end of the input, as shown in Figure 3.19. In other words, input “100” is encoded as “1001.” The simulation setup is the same as stated values earlier in Section 3.4.4, with memristor M1 chosen for simulation, and initial memristance set to 5 k. Tables 3.4 and 3.5 show the effect of encoding for 2-bit input and 3-bit input data, respectively.
b0 b1 b2
b0 b1 b2 b0
b0 b1 b2 b0 b1
3-bit (raw)
Encoded 4bit input
Encoded 5bit input
Figure 3.19 Uniform input encoding of 3-bit input data to 4-bit and 5-bit data Table 3.4 Memristance for 2-bit input with uniform encoding Input
Raw (k)
Encoded value (k)
Encoding scheme
“00” “01” “10” “11”
6.6384 5.0005 4.99955 2.5919
7.2453 5.8998 3.9079 1.0396
“000” “010” “101” “111”
Bold represent additional encoded bits.
Table 3.5 Memristance for 3-bit input with two different uniform encoding schemes
Input
Raw (k)
Encoded to 4-bit (k)
Encoding scheme
Encoded to 5-bit (k)
Encoding scheme
“000” “001” “010” “011” “100” “101” “110” “111”
7.2453 5.8991 5.8997 3.9089 5.8990 3.9078 3.9088 1.0396
7.7431 6.6384 6.6389 5.0004 4.9996 2.5906 2.5919 0.1000
“0000” “0010” “0100” “0110” “1001” “1011” “1101” “1111”
8.1511 7.2451 5.9002 3.9095 5.8968 3.9073 1.0404 0.1000
“00000” “00100” “01001” “01101” “10010” “10110” “11011” “11111”
The background of ReRAM devices
69
Table 3.4 presents the memristance for raw input data and encoded data. Originally, the memristance difference between “01” and “10” is nearly 1 , whereas when encoded to 3 bits, the relative memristance is nearly 1.9 k. As such, larger implementation errors (such as inaccuracy of reference values) and variations (of fabrication, implementation or between different cells) can be tolerated. In [166], a potential of up to 6% physical variation in fabrication process is observed and in [165], 10% memristance variation is assumed. However, given that the minimum distance between the memristance states after the proposed encoding is 19% (whereas it was 0.2% before), a flawless operation despite those variations is expected. Similarly, the effect of encoding on 3-bit data is presented in Table 3.5. Here, we show the memristance when the 3-bit data is encoded to 4-bit, and 5-bit (N.B., in this case first two MSB bits are appended to the end of the data). From Table 3.5, we can see that the memristance difference between the two inputs of “001” and “100” for the raw (3-bit), encoded to 4-bit and encoded to 5-bit data, is, respectively, 0.1 , 1.64 k and 1.4 k. However, if we notice, the relative memristance difference between “001” and “010,” when encoded to 4-bit, is 100), current ReRAMs with high impedance are insignificant, so that the voltages on BLs approximately equal to kVr gon Rs according to (3.9), where k is the number of ReRAMs in low-resistance state (gon ). The underlying assumption is that the number of columns is less than 100 because the ON/OFF ratio is finite. It is obvious that voltages on BLs are all identical. Therefore, the key to obtain the inner product is to set a ladder-type sensing threshold voltages for each column: Vth, j =
(2j + 1)Vr gon Rs 2
xi,1
xi,2 xi,3
(5.3)
φ1,j
φ1,j
φ1,j
φ1,j
φ2,j
φ2,j
φ2,j
φ2,j
φ3,j
φ3,j
φ3,j
φ3,j
Current–voltage converter with comparator Ij
Rfixed
xi,N
φN,j
φN,j SA
SA
φN,j
Vth, j
Vout, j
φN,j SA
SA
O1,0 O1,1 O1,2 Ladder-like Vth in comparators
O1,N-1
Figure 5.1 Parallel digitizing step of ReRAM crossbar in matrix multiplication
XIMA: the in-ReRAM machine learning architecture
101
where Vth, j is the threshold voltage for the jth column. The Oi, j is used to denote the output of column j in ReRAM-crossbar step i after sensing. Therefore, the output of a column is given by 1, j ≤ s O1, j = (5.4) 0, j > s where s is the inner-product result. In other words, the first (N − s) output bits are 0 and the rest s bits are 1 (s 96. From the delay perspective, as shown in Figure 5.5(b), non-distributed ReRAM crossbar is the worst, as it has only one control bus and takes too much time on preparing of computing. Delay of non-distributed ReRAM crossbar grows rapidly while distributed ReRAM crossbar and CMOS-based ASIC implementation maintains on approximately 21 and 70 μs, respectively, as the parallel design. For the energy efficiency, as shown in Figure 5.5(c), both non-distributed and distributed ReRAM crossbars do better as logic accelerator is off at most of time. The discussed architecture also performs the best in energy-delay product (EDP) shown in Figure 5.5(d). Distributed XIMA performs the best among all implementation under different specifications. The EDP is from 0.3 × 10−9 to 2 × 10−9 s · J, which is 60× better than non-distributed ReRAM crossbar and 100× better than CMOS-based ASIC. What is more, hardware performance comparison with varying N is shown in Figure 5.7. Area and energy consumption trend is similar to Figure 5.5. But for the computational delay, the discussed architecture cannot maintain constant delay as shown in Figure 5.5(b) due to larger time required to configure the input, but still the best among the three. Distributed XIMA still achieves better performance than the other two.
106
ReRAM-based machine learning
5.1.2 Coupled ReRAM oscillator network 5.1.2.1 Coupled-ReRAM-oscillator network for L2-norm calculation In a coupled oscillator network, this chapter considers the ReRAM oscillator cell similar to the one shown in Figure 3.15(a) (Rs replaced with a PMOS). By altering the gate
Calculation error rate
0.25 0.2 0.15 0.1 Analog ReRAM 0.05 0
Digitalized ReRAM 0
0.002
0.004
0.006
0.008
0.01
ReRAM error rate
Figure 5.6 Calculation error comparison between multilevel and binary ReRAM 10
CMOS-based ASIC Non-distributed ReRAM crossbar Distributed ReRAM crossbar
1
Delay (μs)
Area (mm2)
1,000 CMOS-based ASIC Non-distributed ReRAM crossbar Distributed ReRAMcrossbar
100
0.1 300
400
500
600 N
700
800
10 300
900
400
500
(a)
600 N
700
800
900
(b)
100
1,000 CMOS-based ASIC Non-distributed ReRAM crossbar Distributed ReRAM crossbar
100
EDP (10–9 s·J)
Energy (μJ)
1,000
10 1 0.1
10 300
400
500
600 N
(c)
700
800
900
0.01 300
CMOS-based ASIC implementation Non-distributed ReRAM crossbar Distributed ReRAM crossbar
400
500
600 N
700
800
900
(d)
Figure 5.7 Hardware performance scalability under different original dimension for (a) area; (b) delay; (c) energy; (d) EDP
XIMA: the in-ReRAM machine learning architecture
107
voltage of the PMOS, one can control the oscillatory frequency as in Figure 3.15(d). Referring to the previous works [209,210], the discussed coupled-ReRAM-oscillator network is shown in Figure 5.8(a). Here, the oscillators are directly coupled with ReRAMs. The oscillatory frequency of each oscillator is determined by multiple input voltages. As mentioned in Figure 3.15(d), the voltage-controlled P-channel metal–oxide–semiconductor (PMOS) can be viewed as a voltage-controlled resistor. Figure 5.8(b) and (c) is the schematic of the oscillators with one input and multiple inputs. After directly coupling, one can have an inverter and a feedback resistor with a diode-connected NMOS. Because of the rectifying property of diode-connected NMOS, these elements guarantee that the current will only pass from the oscillator network to Vout . An RC circuit is design at Vout so that one can measure the DC voltage of Vout . The output voltage indicates the relative synchronization of the coupled oscillators.
5.1.2.2 Performance evaluation The detailed evaluation of ReRAM-based oscillator is shown in Table 5.2. Compared with the design in [209,210], it has power consumption and can be controlled by MemristorMemristor based oscillator (OSC) for coupling
OSC
V21 V22 V23
OSC
Feedback resistor and CMOS inverter
...
V11 V12 V13
V31 V32 V33
(a)
Single input
Vout
OSC
Rout
Cout
Coupled oscillator network with voltage input
VDD
VDD
V1
... V
V2
n
Vout Vout
Vcontrol
Adjustable input patterns (V1, V2,…,Vn)
(b)
(c)
Figure 5.8 (a) ReRAM oscillator network schematic. (b) Oscillator with single input. (c) Oscillator with multiple input
108
ReRAM-based machine learning
multiple voltage inputs easily. As the current through ReRAM-based oscillator is low, the working frequency of the oscillator is lower in some of the compared existing works. As discussed above, Vout is charged by the oscillator network. Therefore, the capacitance of Cout should be large enough to avoid the voltage drop at Vout . At the meantime, large capacitance leads to a large settling time. Figure 5.9 shows the transient response with Cout = 1, 2, 5 and 20 pF, respectively. Obviously Cout = 5 pF is the best configuration because we can capture a stable output at 7.5 μs. In this simulation, the Rout is set as 10 M.
5.2 ReRAM network-based in-memory ML accelerator This section introduces the different computing accelerators based on ReRAM devices. First, the distributed ReRAM-crossbar in-memory architecture (XIMA) will be illustrated in detail. The integration of ReRAM-crossbar-based storage and Table 5.2 Detailed evaluation of ReRAM-based oscillator I max
I min
Power
Frequency
Area
Value
1V
9 μA
3 μA
5.87 μW
6 MHz
0.49 μm2
400 356 311 267 222 178 133 88.9 44.4 0
0
2.5
5 Time (μs) (a)
7.5
0
2.5
5 Time (μs) (c)
7.5
400 356 311 267 222 178 133 88.9 44.4 0
10
V (mV)
V (mV)
400 356 311 267 222 178 133 88.9 44.4 0
V (mV)
V DD
V (mV)
Parameter
10
400 356 311 267 222 178 133 88.9 44.4 0
0
2.5
5 Time (μs) (b)
7.5
10
0
2.5
5 Time (μs) (d)
7.5
10
Figure 5.9 Transient response of ReRAM oscillator network output with (a) Cout = 1 pF; (b) Cout = 2 pF; (c) Cout = 5 pF; (d) Cout = 20 pF
XIMA: the in-ReRAM machine learning architecture
109
computing brings advantages on throughput. After that, two types of 3D CMOSReRAM architectures will be introduced. The major merits of the 3D architecture are less area overhead as well as higher throughput by the 3D interconnection.
5.2.1 Distributed ReRAM-crossbar in-memory architecture Conventionally, the processor and memory are separate components that are connected through I/Os. With limited width and considerable RC delay, the I/Os are considered the bottleneck of system overall throughput. As memory is typically organized in H-tree structure, where all leaves of the tree are data arrays, it is promising to impose in-memory computation with parallelism at this level. In the following section, a distributed ReRAM-crossbar in-memory architecture (XIMA) is discussed. Such an architecture can be used for several ML applications, as presented in Section 7.5.
5.2.1.1 Memory-computing integration As both data and logic units have uniform structure when implemented on ReRAM crossbar, half of the leaves are exploited as logic elements and are paired with data arrays. The discussed architecture is illustrated in Figure 5.10. The distributed local data-logic pairs can form one local data path such that the data can be processed locally in parallel, without the need of being read out to the external processor. Coordinated by the additional controlling unit called in-pair control bus, the in-memory computing is performed in following steps: (1) logic configuration: processor issues the command to configure logic by programming logic ReRAM crossbar into specific pattern according to the functionality required; (2) load operand: processor sends the data address and corresponding address of logic accelerator input; (3) execution: logic accelerator can perform computation based on the configured
General processor Block Data/address/command IO decoder
Bitline BLj
Rij
… …
One-layer ReRAM crossbar
WLi
…
Data
…
Block decoder
… VWL,i
…
Word line
…
Memory
Ij SA
SA
…
SA SA VBL,j
Control bus Ij
Block decoder Local data/logic pair
Multiplelayer ReRAM crossbar
In-memory logic
Rs
Vth, j
VOut, j
Figure 5.10 Overview of distributed in-memory computing architecture on ReRAM crossbar
110
ReRAM-based machine learning
logic and obtain results after several cycles; (4) write-back: computed results are written back to data array directly but not to the external processor. With emphasis on different functionality, the ReRAM crossbars for data storage and logic unit have distinctive interfaces. The data ReRAM crossbar will have only one row activated at a time during read and write operations, and logic ReRAM crossbar; however, one can have all rows activated spontaneously as rows are used to take inputs. As such, the input and output interface of logic crossbar requires analogto-digital (AD) or digital-to-analog (DA) conversions, which could outweigh the benefits gained. Therefore, in this chapter, a conversion-free digital-interfaced logic ReRAM-crossbar design is presented, which uses three layers of ReRAM crossbars to decompose a complex function into several simple operations that a digital crossbar can tackle.
5.2.1.2 Communication protocol and control bus The conventional communication protocol between external processor and memory is composed of store and load action identifier, address that routes to different locations of data arrays and data to be operated. With additional in-memory computation capacity, the presented distributed in-memory computing architecture requires modifications on the current communication protocol. The new communication instructions are proposed in Table 5.3, which is called in-pair control. In-pair control bus needs to execute instructions in Table 5.3. SW (store word) instruction is to write data into ReRAMs in data array or in-memory logic. If the target address is in a data array, it will be a conventional write or result write-back; otherwise it will be logic configuration. LW (load word) instruction performs similar to a conventional read operation. ST (start) instruction means to switch on the logic block for computing after the computation setup has been done. WT (wait) operation is to stop reading from instruction queue during computing. Besides the communication instructions, memory address format is also different from that in the conventional architecture. To specify a byte in the proposed architecture, address includes the following identifier segments. First, the data-logic pair index segment is required, which is taken by block decoders to locate the target
Table 5.3 Protocols between external processor and control bus. Instruction
Operation 1
Operation 2
Action
Function
SW
Address 1
Address 2
Data
Address
LW
Address
—
Store data, configure logic, in-memory results write-back Standard read
ST
Block Index
—
WT
—
—
Address 1 data to Address 2 Store data to address Read data from address Switch logic block on Wait for logic block response
Start in-memory computing Halt while performing in-memory computing
XIMA: the in-ReRAM machine learning architecture
111
data-logic pair. Second, one-bit flag is needed to clarify that whether the target address is in data array or in-memory logic crossbar. Third, if logic accelerator is the target, additional segment has to specify the layer index. Lastly, rest of address segment are row and column indexes in each ReRAM crossbar. An address example for data array and in-memory logic is shown in Figure 5.11. To perform logic operation, the following instructions are required to performed. First, one can store the required input data and ReRAM values with SW operation. Second, an ST instruction will be issued to enable all the columns and rows to perform the logic computing. The WT instruction is also performed to wait for the completion of logic computing. At last, LW instruction is performed to load the data from the output of ReRAM crossbar. Given the new communication protocol between general processor and memory is introduced, one can design the according control bus as shown in Figure 5.11. The control bus is composed of an instruction queue, an instruction decoder, an address decoder and an static random access memory (SRAM) array. As the operation frequency of ReRAM crossbar is slower than that of external processor, instructions issued by the external processor will be stored in the instruction queue first. They are then analyzed by instruction decoder on a first-come-first-serve (FCFS) basis. The address decoder obtains the row and column index from the instruction, and SRAM array is used to store temporary data such as computation results, which are later written back to data array.
5.2.2 3D XIMA 5.2.2.1 3D single-layer CMOS-ReRAM architecture Recent works such as [211–216] have shown that the 3D integration supports heterogeneous stacking as different types of components can be fabricated separately, Module Data array In-memory logic
Block decoder Block index
0 1
Address decoder 0…0 In-layer address Layer index
External processor 2. In data array
4
1
M11 ... ... M1P
... ... ... ... ... ...
2 3
Original matrix in this pair
Row/column to data array
Data path to data array
Logic
Data
MN1... ... MNP
Instruction queue
... ... ... ... ... ...
M11 ... ... M1P
By three-layer ReRAM crossbar logic
MN1... ... MNP
m×m matrix Instruction decoder Address decoder
SRAM array
In-pair CMOS control bus
3. In data array
4. Readout for classification
01010001 Row/column to in-memory logic
Data path to inmemory logic
10110011 …... M×P matrix
IO overhead reduced to M/N
External processor
Figure 5.11 Detailed structure of control bus and communication protocol
112
ReRAM-based machine learning
and layers can be stacked and implemented with different technologies. Therefore, stacking nonvolatile memories on top of microprocessors enables cost-effective heterogeneous integration. Furthermore, works in [217–221] have also shown the feasibility to stack ReRAM on CMOS to achieve smaller area and lower energy consumption. The discussed 3D CMOS-ReRAM accelerator is shown in Figure 5.12(a). This accelerator is composed of a top layer of WLs, a bottom layer of CMOS circuits and vertical connection between both layers by ReRAM. In this architecture, the ReRAM crossbar performs matrix–vector multiplication and also vector addition, as shown in Figure 5.12(b). Here, we use an example Y = X · W to show the mapping scheme of matrix–vector multiplication. The matrix W needs to be stored in ReRAM crossbar by writing the corresponding resistance. To perform the computations, X is converted to WL voltages and one can obtain the output current I denoting Y . One can simply convert the current to voltage and utilize the ADC for digital conversion [147]. The readout circuit of ReRAM crossbar is implemented by the bottom-layer CMOS. In addition, the control signals of the ReRAM crossbar and other operations (pooling, activation) are also realized by CMOS. The 3D single-layer CMOS-ReRAM architecture will be used for tensorized neural network (TNN) in the following chapters. This book will further introduce the detailed mapping and results in Sections 6.2 and 6.2.3.
5.2.2.2 3D multilayer CMOS-ReRAM architecture Besides the 3D single-layer CMOS-ReRAM architecture, this chapter also discusses the 3D multilayer CMOS-ReRAM architecture for online ML applications as well. This 3D multilayer CMOS-ReRAM accelerator with three layers is shown in Figure 5.13(a). This accelerator is composed of a two-layer ReRAM-crossbar and one-layer CMOS circuit. As Figure 5.13(a) shows, Layer 1 of ReRAM crossbar is implemented as a buffer to temporarily store input data to be processed. Layer 2 of ReRAM crossbar performs logic operations such as matrix–vector multiplication and
Step 1: ReRAM configuration Step 2: Image input from processor to wordlines Step 3: Multiplication by ReRAM Step 4: Intermediate data to ReRAM if necessary Step 5: Write-back to processor
ReRAM
WL1 WL2 WL3 WL4
ReRAM
ReRAM
ReRAM
ReRAM
(a)
ReRAM
Input data/ Weight configuration
ReRAM
Intermediate data
Multiplication: Y = XW
Wordline connection (top layer) Vertical connection ReRAM logic CMOS logic (bottom layer)
CMOS processing engine
X1 W 1,1 X2 W
W1,2
W1,3
W1,N
W2,2
W2,3
W2,N
X3 W 3,1 W3,2
W3,3
W3,N
2,1
Wordlines ReRAMs CMOS computation
XM
WM,1 WM,2 WM,3 ADC ADC ADC
Y1
Y2
Y3
WM,N ADC
YN
CMOS-based pooling General processors
Computation output
CMOS-LUT-based activation
(b)
Figure 5.12 (a) 3D CMOS-ReRAM accelerator architecture; (b) ReRAM-based matrix–vector multiplication engine
(b)
Ij
(Layer 1 ) and accelerator (Layer 2)
Rs Vth,j
Operation mapping on CMOS & ReRAM
Register 1
Postnormalization shift Pipeline divider LUT for sigmoid function
Y = sigmoid (X)
ReRAM logic configuration
Key problem reformulation
SA SA VBL, j
VOut,j ReRAM data storage
Integer divider Stage 1 ML Architecture Design
SA
CMOS accelerator
Encoding
XOR
…
SA CMOS dividers and nonlinear function
(a)
Performance evaluation
…
Ij …
General processors
ML analysis
Rij
…
CMOS logic (Layer 3)
WLi …
ReRAM logic (Layer 2)
…
VWL,i …
Wordline
Input data buffer (Layer 1)
Parallel digitizing
Bit line BLj
Matrix–vector multiplication Vector adder Integer divider Stage 2 Register 2
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 –5 –4 –3 –2 –1 0 X
1
2
3
4
5
Figure 5.13 (a) 3D multilayer CMOS-ReRAM accelerator architecture; (b) incremental ML algorithm mapping flow on proposed accelerator
114
ReRAM-based machine learning
also vector addition, as discussed earlier. Note that buffers are designed to separate resistive networks between Layers 1 and 2. The last layer of CMOS contains read-out circuits for ReRAM crossbar and performs as logic accelerators designed for other operations besides matrix–vector multiplication, including pipelined divider, look-up table (LUT) designed for division operation and activation function in ML. Moreover, Figure 5.13(b) shows the workflow for incremental ML based on the proposed architecture. First, the detailed architecture of ML (e.g., number of layers and activation function) is determined based on the accuracy requirements and data characteristics. Second, operations of this ML algorithm are analyzed and reformulated so that all the operations can be accelerated in 3D multilayer CMOS-ReRAM architecture as illustrated in Figure 5.13(a). Furthermore, the bit width operating on ReRAM crossbar is also determined by balancing the accuracy loss and energy saving. Finally, logic operations on ReRAM crossbar and CMOS are configured based on the reformulated operations, energy saving and speedup. Such a 3D multilayer CMOS-ReRAM architecture has advantages in three manifold. First, by utilizing ReRAM crossbar for input data storage, leakage power of memory is largely removed. In a 3D architecture with TSV interconnection, the bandwidth from this layer to next layer is sufficiently large to perform parallel computation. Second, ReRAM crossbar can be configured as computational units for the matrix–vector multiplication with high parallelism and low power. Lastly, with an additional layer of CMOS-ASIC, more complicated tasks such as division and nonlinear mapping can be performed. As a result, the whole training process of ML can be fully mapped to the proposed 3D multilayer CMOS-ReRAM accelerator architecture towards real-time training and testing. The 3D multilayer CMOS-ReRAM architecture will be used for online ML (both learning and inference phases) in the following chapters. We will further introduce the detailed mapping and results in Sections 6.2 and 6.2.3.
Chapter 6
The mapping of machine learning algorithms on XIMA
This chapter presents the mapping of multiple machine learning applications on different resistive random-access memory (ReRAM)-based architectures discussed in Chapter 5. For each application, a detailed mapping strategy, experimental settings and evaluation results are presented here.
6.1 Machine learning algorithms on XIMA First, this chapter presents a detailed mapping of several machine learning applications on distributed in-memory computing architecture (XIMA), including single-layer feedforward neural network (SLFN)-based learning, binary convolutional neural network (BCNN)-based inferences on passive array as well as One Selector One ReRAM (1s1R) array.
6.1.1 SLFN-based learning and inference acceleration The details of SLFN-based learning are presented in [148]. In [148], one maps the SLFN-based learning and inference process according to the algorithm and network in Section 4.2. In this network, matrix–vector multiplication, the major operation in machine learning on ReRAM crossbar is implemented by three steps with detailed mapping strategy for each step. The process of mapping of X A to ReRAM is shown in Figure 4.3 as example. In matrix–vector multiplication, let us assume the output matrix is Y. For better understanding, we replace X with X and A with , and the equation becomes Y = X. For every element in Y, it follows yij =
N
xik ϕkj
(6.1)
k=1
where x and ϕ are the elements in X and , respectively. The calculation of yij by four-step ReRAM crossbar is shown in the following subsections.
6.1.1.1 Step 1. Parallel digitizing The first step parallel digitizing requires an N × N ReRAM crossbar. The idea is to split the matrix–vector multiplication to multiple inner-product operations of two
116
ReRAM-based machine learning
vectors. Each inner product is produced ReRAM crossbar as shown in Figure 6.1(a). x is set as crossbar input and ϕ is written in ReRAM cells. In N × N ReRAM crossbar, resistance of ReRAMs in each column is the same, but V th among columns are different. We have (2j + 1)Vr Rfixed Vth, j = (6.2) 2Ron which are ladder-like threshold voltages to identify the result. If the inner-product result is s, the output of columns will be 1, j ≤ s O1, j = (6.3) 0, j > s
6.1.1.2 Step 2. XOR The second step is exclusive OR (XOR) to identify the index of s where O1,s = 1 and O1,s+1 = 0, so the operation is O1,s ⊕ O1,s+1 . According to (6.3), the case O1, j = 0 and O1, j+1 = 1 is impossible, so the XOR operation can be converted to O1,s O1,s+1 and O1,s + O1,s+1 . The mapping of ReRAM-crossbar input and resistance in this step is shown in Figure 6.1(b), and threshold voltage configuration is Vth, j =
Vr Rfixed 2Ron
(6.4)
Therefore, only the output of sth column is 1 and the others are 0.
6.1.1.3 Step 3. Encoding The third step takes the output of XOR step and produces s in binary format as an encoder. The input of ReRAM crossbar is from the second step and only one input is 1, so that according binary format is stored in corresponding row, as shown in N O1,0 = ∑Xi,kφk,j k=1
Xi,1 φ1,j Xi,2 φ2,j Xi,3 φ3,j
O2,0 = O1,0+O1,1
φ1,j
φ1,j
φ2,j
φ2,j
φ3,j
φ3,j
Xi,4 φ4,j
φ4,j
φ4,j
Xi,N φN,j O1,0
φN,j O1,1
φN,j
(a)
O1,N-1
O1,0 O1,1 O1,1 O1,2
O1,N-1 O2,0 (b)
O3 = Binary(s), if O2,s = 1 Binary O2,0 of 0 (00) Binary O2,1 of 1 (01) 0 or 1 Binary O2,2 of 2 (10) Binary 1 O2,3 of 3 (10) Binary of N
O2,N O2,1
O2,N-1
O3,0 (c)
O3,n-1
0
O3,n (d)
Figure 6.1 Detailed mapping for digitalized matrix–vector multiplication: (a) parallel digitizing; (b) XOR 2; (c) encoding; (d) legend of ReRAM mapping
The mapping of machine learning algorithms on XIMA 117 Figure 6.1(c). Encoding step needs an N × n ReRAM crossbar, where n = log2 N is the number of bits in order to represent N in binary format. The thresholds for the encoding step are set following (6.4) as well.
6.1.1.4 Step 4. Adding and shifting for inner-product result The output of encoding step is in binary format, but some processes are needed to obtain the final inner-product result (Figure 6.2). Adder and shifter are designed to complete this process as shown in Figure 6.1(d). Let us consider the original data are 8-bit and data dimension is 512, the workload of adder is 512 without any acceleration for the purpose of illustration. With four-step ReRAM-crossbar accelerator as preprocessing, the workload of adder can be significantly reduced to 9 (log2 512).
6.1.2 BCNN-based inference acceleration on passive array The details of this section have been presented in [222]. As indicated in previous works, ReRAM crossbar can be applied to perform logic operations, especially the dot-product-based convolutions. Previous mapping of convolutional neural network (CNN) are mainly based on the traditional analog ReRAM crossbar [145,150]. The main limitation is that there can exist a huge nonuniform analog resistance for undetermined states in the analog computation, which can result in convolution error. Here, we will map the bitwise CNN-based inference process according to the algorithm and network in Section 4.3.3. This section focuses on the mapping of all the bitwise CNN operations on the passive array-based XIMA such as convolution, batch normalization, pooling and activation functions.
6.1.2.1 Mapping bitwise convolution According to (6.5), the bitwise convolution can be split into several XNOR and bitcount results of two vectors. To implement (6.5), we can use two AND operations for abk−1 and wkb as well as their complements. Therefore, the bitwise convolution on ReRAM crossbar can be shown as follows: sk =
N
wkb
⊗
abk−1
=
i=1
N
(wkb · abk−1 + wkb · abk−1 )
(6.5)
i=1
Result of ReRAM encoding 1st bit result
n bits
2nd bit result
110010001 010110101
Adder n bits
Shift right
Control +
Product
Final result
100101110
8th bit result
Inner-product result
Figure 6.2 ReRAM-based inner-product operation
118
ReRAM-based machine learning
The mapping of the bitwise convolution is shown in Figure 6.3(a). It requires a 2N × N ReRAM crossbar, where N is the number of element in the vector. All columns are configured with the same elements that correspond to one column in binary weight wkb of the neural network, and the wordline (WL) voltages are determined by the binary input abk−1 . Due to the large ratio between Roff and Ron , the current through the bitline (BL) is approximately equal to skRVonr Rs , where sk is the inner-product result in (6.6). Since the current of all BLs is identical, the ladder-like threshold voltages Vth, j are set as follows: (2j + 1)Vr Rs Vth, j = (6.6) 2Ron where Vth, j is the threshold voltage for the jth column. If one uses s to denote the output array, and s ( j) to denote the output of column j in ReRAM crossbar, we can have 1, j < sk sk ( j) = (6.7) 0, j ≥ sk In this case, the inner-product results sk can be recognized that the first (N − sk ) output bits are 0 and the rest sk bits are 1. The relation between sk and sk can be expressed as sk = g(sk ). As described in (6.5), each binary weight vector wk performs bitwise convolution with several input features. As a result, each logic block in Figure 6.2 stores a binary vector wk , while the control bus transmits the input feature sequentially. In this case, bitwise convolution can be performed in parallel in separated logic blocks.
6.1.2.2 Mapping bitwise batch normalization Bitwise batch normalization requires two digital ReRAM crossbars in the implementation. In the first ReRAM crossbar, it performs the XOR operation on adjacent bits of the output of bitwise convolution. It can be expressed as 1, j = sk sk ( j) = (6.8) 0, j = sk After that, the second ReRAM crossbar builds a look-up table (LUT). Since μ, σ 2 , γ and β are all fixed in the inference stage, (6.8) can be rewritten as ak = f (sk )
(6.9)
where f (·) represents the LUT. As a result, the LUT is stored in the second ReRAM crossbar according to the parameters μ, σ 2 , γ and β. As described in (6.8), only the sk th row of the LUT is selected, so the batch normalization result can be directly read out. The threshold voltage of both ReRAM crossbars is Vr Rs Vth = (6.10) 2Ron To have a better illustration, Figure 6.3(b) shows the detailed mapping and Table 6.1 shows the values to store in the second ReRAM crossbar when μ = 2.5, σ 2 = 5, γ = 1 and β = 0 referred to the IEEE-754 standard.
ak(x) = (sk(x)-μk(x))/σk2(x)+βk(x)ak(x) = f(sk(x))
wk(N-1)
ak-1(2)
1
0
0
VrRs 2Ron
0
0
1
0
VrRs 2Ron
0
0
0
0
VrRs 2Ron
f (2)
0
0
1
VrRs 2Ron
f (3)
0
wk(N-1)
ak-1(1) wk(N-1)
wk(N-1)
wk(N-1)
wk(N-1)
wk(N-2)
wk(N-2)
wk(N-2)
wk(N-2)
wk(1)
wk(1)
wk(1)
wk(1)
ak-1(N-1)
Ladder-like thresholds
3VrRs 2Ron
5VrRs 2Ron
S′(1)
S′(2) 1
(2N +1)VrRs 2Ron
7VrRs 2Ron
S′(3) 1
S′(1)
S′(N) 1
1
1
S′(2)
S′(2)
S′(N)
1
1
1
f(1)
f(N) Identical thresholds
Bit information
Vr Rs 2Ron
Vr Rs 2Ron 1
(a)
(b)
a′k(i, j) = Max(ak(Mi+k, Mj+l))
1
Signbit(ak( Signbit(ak( Signbit(ak( Signbit(ak Mi+k, Mj+l)) Mi+2, Mj)) Mi+1, Mj)) (Mi, Mj))
ak-1(1) 1-bit ANDwk(N-1)
Vr Rs 2Ron 1
Vr Rs 2Ron
Vr Rs 2Ron 1
1
(c)
Figure 6.3 Digital ReRAM-crossbar mapping for (a) bitwise convolution, (b) bitwise batch normalization and (c) bitwise max-pooling
120
ReRAM-based machine learning Table 6.1 Binary format stored in ReRAM crossbar BinConv output
Batch normalization
Floating binary format
0 1 ··· 7
–0.5 –0.3 ··· 0.9
10111111000000000000000000000000 10111110100110011001100110011010 ··· 00111111011001100110011001100110
6.1.2.3 Mapping bitwise pooling and binarization According to (6.9), one can perform the binarization first and then perform maxpooling. The bitwise activation in (6.8) can be achieved by selecting the sign-bit of the binary format output of Figure 6.3(b). In max-pooling, the output is 0 only when all the numbers in the pooling region are negative. As a result, one can add all the complementary sign-bit in the pooling region. If the result is not 0, it indicates that at least one positive number is available in the pooling region, resulting in the pooling result 1. In summary, the max-pooling in (6.5) can be rewritten as ak (i, j)
1, if Sign{ak (Mi + k, Mj + l)} > 0 = 0, otherwise
(6.11)
As a result, both of the bitwise max-pooling and binarization can be implemented by the addition operation performed on the digital ReRAM crossbar with 1 column, as shown in Figure 6.3(c).
6.1.2.4 Summary of mapping bitwise CNN As a summary, the four operations of the bitwise CNN in Figure 4.7(b) can be fully mapped onto the digital ReRAM crossbar. All the threshold voltages are fixed even if the parameters of bitwise CNN are changed. Although these operations are implemented in different ReRAM-crossbar array, the input/output formats of them are compatible so that one can directly connect them as shown in the logic block in Figure 6.3 with pipeline used. The pipeline design is based on Table 6.2, so that each stage implements a layer. CONV-2, CONV-4 and CONV-6 are the stages which require the most ReRAM cells and computation time. As a result, more logic blocks are assigned to these steps to relieve the critical path. In our simulation, half of the digitized ReRAM crossbars are used to perform these three layers. Moreover, since the output feature from Figure 6.3(c) is binary, the area overhead and energy consumption of the data storage can be significantly reduced. In addition, because layers in Table 6.2 are implemented in different data-logic pairs, the volume of data transmitted is also decreased.
The mapping of machine learning algorithms on XIMA
121
Table 6.2 Bitwise CNN configuration of CIFAR-10 data set Name
Filter/weight
Output size
CONV-1 CONV-2 MP-2 CONV-3 CONV-4 MP-4 CONV-5 CONV-6 MP-6 FC-1 FC-2 FC-3
3 × 3 × 3 × 128 3 × 3 × 128 × 128 2×2 3 × 3 × 128 × 256 3 × 3 × 256 × 256 2×2 3 × 3 × 256 × 512 3 × 3 × 512 × 512 2×2 8, 192 × 1, 024 1, 024 × 1, 024 1, 024 × 10
128 × 32 × 32 128 × 32 × 32 128 × 16 × 16 256 × 16 × 16 256 × 16 × 16 256 × 8 × 8 512 × 8 × 8 512 × 8 × 8 512 × 4 × 4 1, 024 1, 024 10
6.1.3 BCNN-based inference acceleration on 1S1R array In this work, we will present mappings of bitwise CNN operations on the sneak-pathfree digital 1S1R ReRAM crossbar. The mapping detail is a little different compared to the mapping on passive array. Different from the signed realization in previous works [149,223], we will use unsigned bitwise operation in the mapping.
6.1.3.1 Mapping unsigned bitwise convolution As described in Section 6.1.2, dot-product operation is the majority in bitwise convolution. Suppose that a single dot-product operation has N numbers, the basic operation can be expressed as sx,y =
N
ax,i · wi,y
(6.12)
i=1
where ax , i, wi,y ∈ {−1, +1}. To implement the negative weights, [149,223] use two ReRAM crossbars to represent positive and negative weights, respectively. However, this solution requires a subtraction operation for every complementary BL output. It is obvious that the subtraction implementation leads to extra CMOS circuit with more area and energy consumption. To avoid this problem, one can apply the conversion by XNOR operation ( ) in {0, 1} introduced in [189,224], as shown in Table 6.3. Since a w = a · w + a · w , one can directly map a , a as WL inputs, and w , w as ReRAM conductances. We can use s to denote the new bitwise convolution result as sx,y
=
N i=1
ax,i
wi,y
=
N i=1
(ax,i · wi,y + ax,i · wx,i )
(6.13)
122
ReRAM-based machine learning Table 6.3 Conversion from signed bitwise convolution to unsigned bitwise XNOR a
w
Multiplication
a
w
XNOR
+1 +1 −1 −1
+1 −1 +1 −1
+1 −1 −1 +1
1 1 0 0
1 0 1 0
1 0 0 1
Since s must be an even integer and s ∈ [ − N , N ] while s ∈ [0, N ], the relation can be summarized as s = 2s − N . Although this mapping strategy requires the same number of ReRAM devices as the one in [149,223], it only needs an inverter for each input bit.
6.1.3.2 Mapping batch normalization, pooling and binarization Section 4.3.3 has explained these three steps with detailed equations, but it is simpler to map them on ReRAM crossbar. We can notice that the order of binarization and pooling does not affect the final output. Therefore, by combining batch normalization √ and binarization, we can use c = μ − β σ 2 /γ from BatchNorm (parameters for the {0, 1} convolution fashion) as the threshold to compare with the bitwise convolution result. Although the corresponding voltage of c is generated by limited-bit DAC, it is still more energy-efficient than the ADC for analog ReRAM crossbar. The comparison process is binarization with only 0 or 1 output. These two steps can be summarized as √ N 2 1, i=1 ax,i wi,y ≥ μ − β σ /γ Out = (6.14) √ N 2 0, i=1 ax,i wi,y < μ − β σ /γ As for pooling operation, we can put a single-output ReRAM crossbar and conr Rs figure the comparator threshold as 2(RsV+R to detect if any binarization result in the LRS ) pooling region is 1. Figure 6.4 shows the summary of bitwise CNN mapping on the proposed digital ReRAM crossbar. Since the output of each BL is also in {0, 1} fashion, it can be directly used for the binary input feature of the next layer.
6.1.4 L2-norm gradient-based learning and inference acceleration The details of this section have been presented in [225]. This section illustrates mapping of the multilayer feedforward neural network-based learning and inference processes according to the algorithm and network in Section 4.2.2. This section focuses on the mapping of all the matrix–vector multiplication operations and L2-norm calculation on the passive array and coupled oscillator-based XIMA.
The mapping of machine learning algorithms on XIMA
123
Parallel bitwise convolution s(N) = a(x)Ĵw(x), where a(x), w(x) ę{0, 1}
a′1 w′11
w′12
w′13
w′11
w′12
w′13
w′1M …
a′1 w′1M
Binary input feature
… a′2 w′21
w′2M
w′23
w′22
… …
…
w′N2
w′N3
c1
c2
c3
…
… w′N1
a′N
w′NM …
cM
Pooling output for the next layer Max (s′(1), s′(2), ...) LRS
LRS
canalog (12.5, 13.5, etc)
Rs cdigital
c Out s′
DAC c = μ - β σ2 γ
Configured by parameters in batch normalizaion
I
Binarization Out = 1, if s′ϊc Out = 0, if s′ ≤ c
Figure 6.4 Mapping of bitwise CNN on sneak-path-free digital ReRAM crossbar
124
ReRAM-based machine learning
6.1.4.1 Mapping matrix–vector multiplication on ReRAM-crossbar network With the introduction of binary-memristor-crossbar network, one can perform matrix– vector multiplication [147]. It has much better nonuniformity under process variation when compared to the multiple-level-valued analog memristor crossbar. As discussed in Section 6.1.2, the key operations in machine learning (feature extraction and convolution) requires a large number of matrix–vector multiplication. As discussed earlier, the key operations (6.13) and (6.14) requires a large number of matrix–vector multiplication. For instance, consider Y = HW as an example for mapping multiplication on crossbar. For every element in Y, it follows, here, we take the operation of Y = HW as an example for mapping the multiplication on the crossbar. For every element in Y, it follows yij =
N
hik wkj
(6.15)
k=1
where h and w are the elements in H and W, respectively. For the crossbar structure in Figure 6.4, the current through each BL follows Ij =
m Vwl,i Rfixed i=1
Rij
(6.16)
As a result, we can map hik and wkj on Vwl,i and Rij , respectively. By measuring the current through BL with binary digitization, one can derive the result yij in the digital domain. The detailed operation can be referred to the work in [147]. Note that for real numbers, one needs only a few more shifted additions to complete the final process of matrix–vector multiplication with memristor crossbar. Figure 6.5 shows that we only need four-shifted addition with 8-bit precision to calculate the final inner product. In addition, Sigmoid function can be implemented in memristor crossbar with LUT.
6.1.4.2 Mapping L2-norm calculation on coupled ReRAM oscillator network In a coupled oscillator network, we use the memristor oscillator cell similar to that in Figure 3.15(a) (Rs replaced with a PMOS). By altering the gate voltage of the PMOS, we can control the oscillatory frequency as in Figure 3.15(d). Referring to previous works [209,210], the proposed coupled-memristor-oscillator network is shown in Figure 5.8(a). Here, oscillators are directly coupled with memristors. The oscillatory frequency of each oscillator is determined by multiple input voltages. As mentioned in Figure 3.15(d), the voltage-controlled PMOS can be viewed as a voltage-controlled resistor. Figures 5.8(b) and 5.8(c) are the schematic of oscillators with single input and multiple inputs. After directly coupling, one can have an inverter and a feedback resistor with a diode-connected NMOS. These elements guarantee that the current will only pass from the oscillator network to Vout . An RC circuit is designed at Vout so that we can measure the DC voltage of Vout . The output voltage indicates the relative synchronization of the coupled oscillators.
The mapping of machine learning algorithms on XIMA Parallel digitizing hi,1 hi,2 hi,3
hi,N
w1,j
w1,j
w1,j
w2,j
w2,j
w2,j
w3,j
w3,j
w3,j
wN,j
wN,j
O1,0 O1,1
wN,j
XOR O1,0
1 O1,1 1 O1,1 0 O1,2 0 O1,N-1
O1,N-1
125
Encoding
0
0
O2,0
0
0
O2,1
Binary of 2
1
0
O2,2
Binary of 3
1
0
0 0 1 O2,0 O2,1 O2,N-1
Binary of 1
Binary of N
O2,N-1 O3,0 O3,1 O3,2
O3,n
Binary format result
Output of memristor crossbar network
(a) Inner product of the 5th LSB 1bit result Inner product of the 4th LSB
+
(b)
Inner-product result of 16 8-bit numbers
Figure 6.5 (a) Crossbar network for multiplication; (b) shifted additions
To perform (4.14) on oscillator network, we only need to use two inputs for each oscillator in Figure 5.8(c). To be more specific, one can have a mapping function as follows: Vik1 = Vbias + αtik
Vik2 = Vbias − αyik
(6.17)
where α is a scale factor to link the oscillator input voltage, and tik and yik are the elements in matrices T and Y, respectively. As mentioned in [209,210], the coupled oscillator network can work as a degreeof-match (DOM) circuit to calculate L2-norm by fitting Vout to 2nd polynomial when there are two inputs for each oscillator in Figure 5.8(c). Therefore, the output voltage follows Vout = Vmin + η||T − Y||22 = Vmin + η
M N
(tik − yik )2
(6.18)
k=1 i=1
where η is a fitting parameter, and Vmin is the output voltage when ||T − Y||22 = 0. As such, one can use the coupled-memristor-oscillator network to the L2-norm error function in machine learning. The L2-norm result is obtained by curve fitting but not accurate computing. Compared to other implementations, such an oscillator network can complete the operation much faster because the square operation (tik − yik )2 can be performed simultaneously by the oscillator circuit. Moreover, the accumulation is also automatically achieved by the coupled network.
126
ReRAM-based machine learning
6.1.4.3 Mapping flow of multilayer neural network on ReRAM network Based on the coupled-memristor-oscillator network and binary-memristor-crossbar network, we can accelerate direct-gradient-based L2-norm minimization for machine learning as shown in Figure 6.6. For the model part, we use two types: memristor oscillator and memristor crossbar. We find the parameter in oscillator network to calculate L2-norm, and map on crossbar network to calculate digitized matrix–vector multiplication. Afterwards, one can generate the initial weights of A and W to calculate the gradient δE. The best direction to minimize E can be derived. This process will be repeated several times till the minimization is reached only when E k − E k−1 < threshold.
6.1.5 Experimental evaluation of machine learning algorithms on XIMA architecture We will show the detailed mapping of several machine learning applications on XIMA, including SLFN-based learning, BCNN-based inferences on passive array as well as 1S1R array.
Memristor Model
Memristor oscillator
Direct-gradientbased L2-norm minimization Forming neuron network Weights update Matrix–vector multiplication
Random generated weights
Memristor crossbar
Passing forward network Parameters tuning for L2-norm
L2-norm on oscillator network Alternative L2 computation
Device tuning and operation initialization
Direct gradient computation Update each weight W
Complete L2-norm minimization
E minimized for every w?
Figure 6.6 Mapping flow of machine learning on memristor network
The mapping of machine learning algorithms on XIMA
127
6.1.5.1 SLFN-based learning and inference acceleration Experiment settings The hardware evaluation platform is implemented on a computer server with 4.0 GHz core and 16.0 GB memory. Feature extraction is implemented by general processor, CMOS-based application-specific integrated circuit (ASIC), non-distributed and distributed in-memory computing based on digitalized ReRAM crossbar, respectively. For the ReRAM-crossbar design evaluation, the resistance of ReRAM is set as 1 k
and 1 M as on-state and off-state resistance, respectively, according to [206]. The general processor solution is performed by MATLAB® on a 4.0 GHz desktop. For CMOS-ASIC implementation, we implement it by Verilog and synthesize with CMOS 65-nm low-power process design kit (PDK). For ReRAM-crossbar-based solution, we verify the function in circuit level with SPICE tool NVM-SPICE [63]. By analyzing the machine learning algorithm, we obtain the basic operations and the number of ReRAM-crossbar logic units required. The working frequency of general processor implementation is 4.0 GHz while the CMOS-ASIC feature extraction design frequency is 1.0 GHz. For in-memory computing based on the proposed ReRAM crossbar, write voltage Vw is set as 0.8 V and read voltage Vr is set as 0.1 V as well as duration time of 5 ns. In addition, the analog computation on ReRAM crossbar is performed for comparison based on design in [207]. In the following, we will show the performance of matrix–vector multiplication on ReRAM crossbar first. A scalability study is introduced to show the area, energy and computation delay with different matrix sizes. Afterwards, the evaluation of face recognition on in-memory architecture is presented. Finally, we will illustrate the object classification on three-dimensional (3D) CMOS-ReRAM architecture. Performance of different bid-width configurations will be also shown. In addition, the 3D CMOS-ReRAM solution will be compared with CMOS-ASIC as well as Graphics Processing Unit (GPU) implementation.
Performance evaluation In this work, we implement the face recognition application on the in-memory architecture. We will analyze the computation complexity of face recognition first, and then evaluate the performance. In the experiment, 200 face images of 13 people are selected from [226], with scaled image size 262 of each image X. In principal component analysis (PCA), feature size of image is further reduced to 128 by multiplying the matrix R. The number of hidden node L and classes m are 160 and 13, respectively. Based on the experimental settings, computation complexity is analyzed with results shown in Figure 6.7. Here 82% of computations are multiplication in output weight calculation, which is the most time-consuming procedure in neural network. Time consumption of each process in neural network is introduced in Figure 6.7(b). Since processes except activation function involve matrix–vector multiplication, we extracted this operation in the whole algorithm and found that 64.62% of time is consumed in matrix–vector multiplication, as shown in Figure 6.7(c).
128
ReRAM-based machine learning
We implement the face recognition applications by the SLFN on XIMA. In Table 6.4, general performance comparisons among MATLAB, CMOS-ASIC and ReRAM-crossbar accelerator are introduced, and the acceleration of each procedure as the formula described in Section 7.2 is also addressed. Among three implementations, ReRAM-crossbar architecture performs the best in area, energy and speed. Compared to MATLAB implementation, it achieves 32.84× speedup, 210.69× energy-saving and almost four-magnitude area-saving. We also design a CMOS-ASIC implementation with similar structure as ReRAM crossbar with better performance compared to
Add/sub 4% Division 14%
Output weight 36.38%
Multiplication 82% (a)
Random Output weight 6.55% input weight 6.91% preH 34.79% Activation Multiplication 21.92% 64.62%
Activation 21.92% (b)
(c)
Figure 6.7 (a) Time-consumption breakdown for output weight calculation; (b) neural network training computation effort analysis; (c) multiplication analysis for neural network training (N = 200, n = 128, L = 160 and m = 13) Table 6.4 Face recognition performance comparison under different software and hardware implementations Implementation
Frequency Area 2
Power
Type
84 W
PCA 1.56 ms Input layer 0.98 ms L2-norm 92.5 ms Output layer 0.1 ms Overall 95.14 ms PCA 195 μs Input layer 121.9 μs L2-norm 12.26 ms Output layer 12.4 μs Overall 12.59 ms Improvement 7.56× PCA 38.4 μs Input layer 24 μs L2-norm 2.83 ms Output layer 2.4 μs Overall 2.90 ms Improvement 32.84×
CPU (MATLAB)
4.0 GHz
177 mm
CMOS-ASIC
1.0 GHz
5.64 mm2
39.41 W
Distributed in-memory ReRAMcrossbar architecture
200 MHz
0.11 mm2
13.1 W
Time
Energy 131 mJ 82 mJ 7.77 J 8.40 mJ 7.99 J 7.68 mJ 4.80 mJ 483 mJ 0.49 mJ 496 mJ 16.11× 0.5 mJ 0.3 mJ 37.1 mJ 30 μJ 37.93 mJ 210.69×
The mapping of machine learning algorithms on XIMA
129
MATLAB. ReRAM-crossbar architecture is 4.34× speedup, 13.08× energy-saving and 51.3× area-saving compared to CMOS-ASIC. The performance comparison is quite different from Table 5.1, because we applied different designs (ReRAM-crossbar size) is this two experiments according to the dimension of matrices. The result of face recognition is shown in Figure 6.8. Five training classes are provided as an example with three test cases. Each test face will be recognized as the class with the largest score (prediction result as the index of Max(Y), marked in red color). In this example, Case 1 is identified as Class 1, while Cases 2 and 3 are classified into Classes 3 and 5, respectively.
6.1.5.2 L2-norm gradient-based learning and inference acceleration Experiment settings The hardware evaluation platform is implemented on a computer server with 4.0 GHz core and 16.0 GB RAM. For ReRAM-based oscillator network, we use the VerilogA model from [153] and device model from [151] as the ReRAM before forming process and simulate the network in Cadence Spectre. For ReRAM-based crossbar network, the resistance of ReRAM is set as 1 k and 1 M as on-state and off-state resistance, respectively, according to [206]. NVM-SPICE [63] is used for ReRAM-crossbar network verification. For comparison propose, we also implement the L2-norm gradient-based machine learning on general processor. The working frequency of general processor implementation is 3.46 GHz with 64.0 GB RAM. In addition, University of California Irvine (UCI) benchmark data are selected for comparisons [184]. Write voltage Vw is set as 0.8 V and read voltage Vr is set as 0.1 V as well as duration time of 5 ns.
Performance evaluation To evaluate the proposed ReRAM network for machine learning, we use Iris in UCI data set as benchmark. First, we analyze the computation complexity in L2-norm-gradient-based machine learning. Figure 6.9(a) shows the time consumption of CMOS-CPU-based implementation. In simulation, we need to optimize the value
Class 1
Test cases/ classes
Test case 1 (Y) Test case 2 (Y) Test case 3 (Y)
Class 2
Class 3
Class 4
Class 1
1.518062
–0.79108
–0.58029
Class 2 Class 3
–0.29803
–0.87155
–0.24397
–0.95114
0.793256
–0.35867
Class 4
–0.65597
–0.44714
–0.70879
Class 5
–0.65955
0.262689
0.872497
Class 5
Figure 6.8 Training samples and prediction value Y from (4.7) for face recognitions
130
ReRAM-based machine learning Multiplication
L2-norm calculation
Others
1.96% 26.30%
26.41%
71.63%
29.30%
44.40% (b)
(a) 0.44%
0.002% 8.708%
44.46% 55.10% 91.290% (c)
(d)
Figure 6.9 Time-consumption proportion of Iris data set: (a) without ReRAM network; (b) with ReRAM network, and energy consumption proportion of Mushroom data set; (c) without ReRAM network; (d) with ReRAM network of 32 weights to minimize the L2-norm. 71.63% computation time is consumed by matrix–vector multiplication, and 26.41% time is consumed by L2-norm calculation. Therefore, the total computation time can be significantly reduced by accelerating these two operations. To evaluate the ReRAM network performance in Iris data set, we use the proposed ReRAM network to address matrix–vector multiplication and L2-norm calculation. For the other operations (1.96% in Figure 6.9(a)), we still use CPU for calculation. As there are 32 weights to be updated in the benchmark, we design 32 ReRAMoscillator networks to compute E for each w simultaneously. In the experiment, we need to repeat this procedure for 7 times to derive the minimum of L2-norm. The time consumption with ReRAM network is shown in Figure 6.9(b). We can see that L2-norm becomes the dominant time-consuming operation. In addition, the matrix– vector multiplication and L2-norm take only 3× as the other operations, which only take little time in machine learning. Figure 6.9(c) and (d) shows the energy efficiency achieved by the proposed ReRAM network. The energy consumption of machine learning on ReRAM network is significantly reduced. Note that each ReRAM-oscillator-network requires 7.5 μs to perform the one-time L2-norm calculation with 5.87 μW according to Table 5.2.
The mapping of machine learning algorithms on XIMA
131
Note that matrix–vector multiplication and L2-norm calculation are dominant in CMOS-CPU-based implementation. For Mushroom data set, they take over 99% of energy consumption. When we map these operations on the ReRAM network, their energy consumption is only 8.7%. We use four benchmarks in UCI data set [184] to measure the speedup and energy-saving, as shown in Table 6.5. Using the proposed ReRAM network, we can achieve 10× speedup and more than 50× energy-saving. Table 6.5 Time and energy comparison on UCI data set
Benchmark
Training sample
Hidden nodes
Classes
Specifications
CPU
CPU with ReRAM
Iris
114
32
3
Mushroom
7333
40
2
Wine
160
40
3
Liver
317
232
2
Shuttle
14,500
170
7
Glass
192
105
7
Time Multiplication L2-norm Total energy∗ Speedup Energy-saving Time Multiplication L2-norm Total energy∗ Speedup Energy-saving Time Multiplication L2-norm Total energy∗ Speedup Energy-saving Time Multiplication L2-norm Total energy∗ Speedup Energy-saving Time Multiplication L2-norm Total energy∗ Speedup Energy-saving Time Multiplication L2-norm Total energy∗ Speedup Energy-saving
0.0159 s 1.482 J 0.546 J 2.007 J – – 0.0574 s 3.315 J 4.108 J 7.462 J – – 0.0409 s 3.562 J 1.69 J 5.317 J – – 0.521 s 13.77 J 0.546 J 53.86 J – – 3.35 s 11.27 J 0.546 J 424.27 J – – 0.4 s 27.64 J 24.22 J 52 J – –
0.0012 s 1.39 mJ 0.1 μJ 0.042 J 13.47× 49.21× 0.005 s 3.0 mJ 7.584 μJ 0.0357 J 11.48× 209.02× 0.0029 s 3.3 mJ 0.3 μJ 0.0639 J 13.96× 83.21× 0.0558 s 12.9 mJ 0.1 μJ 0.12 J 9.34× 564.42× 0.411 s 10.56 mJ 0.1 μJ 0.0608 J 8.15× 7162.83× 0.0304 s 25.90 mJ 4.47 μJ 0.109 J 13.16× 477.06×
∗ The total energy includes multiplication, L2-norm and other operations (e.g., gradient calculation) performed in CPU.
132
ReRAM-based machine learning
Note that the improvement depends on the training size of benchmark. For example, the L2-norm calculation is the dominant operation in Mushroom and Liver so that we can save over 99.5% energy consumption when using the proposed ReRAM network. For Iris and Wine, matrix–vector multiplication is dominant so that the energy-saving is below 99%.
6.1.5.3 BCNN-based inference acceleration on passive array Experimental settings Baselines In the simulation, we have implemented different baselines for comparison using both MNIST and CIFAR-10 benchmarks. The detail of each baseline is listed as follows: CPU: The BNN simulation is run in Matconvnet [227] on a computer server with 3.46 GHz core and 64.0 GB RAM. The BNN network is referred to Table 6.2. Design of digital ReRAM: The ReRAM device model for BNN is based on [63,64,139] with the resistance of ReRAM set as 0.5M and 5M as on-state and offstate, respectively, with working frequency of 200 MHz. Sense amplifier is based on the design of [228]. The BNN network is also referred to Table 6.2. Others: GPU-based [188] bitwise CNN implementations and field programmable gate array (FPGA)-based [229], CMOS-ASIC-based [230] and analog-ReRAMbased [146] conventional CNN implementations are selected for performance comparisons as well. Network The overall network architecture of the bitwise CNN (called BNN) is shown in Table 6.2. It has six binary convolutional (BinConv) layers, three max-pooling (MP) layers and three fully connected (FC) layers. It takes a 32 × 32 RGB image as the input of the first layer. We use 128 sets of binary filters, and each set contains three binary filters to process the data from R, G and B channels, receptively. The width and height of each BinConv layer is fixed to 3 × 3 with stride of 1 and zero padding of 1. The BinConv layer performs the bitwise convolution between input feature maps and weights followed by the bitwise batch normalization and the binary activation. Note that the computation for FC layers can be treated as a convolution with stride of 0 and zero padding of 0. Thus, we refer the matrix multiplication in FC layers to convolution as well. As shown in Table 6.2, two cascaded convolutional layers form a convolutional block with equivalent 5 × 5 convolution window. This configuration provides more powerful representation capacity with less amount of weights compared with the direct 5 × 5 implementation. Meanwhile, binary batch normalization is applied after the convolution to accelerate and stabilize the training. We adopt the binary activation to each BinConv layer and FC layer with the exception for the last FC layer. The output of the last FC layer is fed into the softmax layer [231] without binarization to generate a probabilistic distribution of ten classes. Besides CIFAR-10, we also show the experiment on MNIST data set [232], with detailed parameters in Table 6.6, which is pretrained on a server. Compared to parameters of CIFAR-10, we use conventional 5 × 5 convolutional cores in MNIST.
The mapping of machine learning algorithms on XIMA
133
Performance evaluation We first show accuracy comparison between the analog ReRAM and the digitized ReRAM under device variation. We then show accuracy comparison between the conventional CNN with direct-truncated prevision and the proposed bitwise CNN. Various benchmarks such as CIFAR-10 [233] and MNIST [232] are used here. Error under device variation In the previous discussion, the digitized ReRAM crossbar has better programming accuracy than the analog ReRAM crossbar. Figure 6.10 shows a 200-time Monte-Carlo simulation of a single ReRAM device programming process with different write voltage Vw , where the voltage amplitude is under Gaussian distribution on the(3σ = 3%Vw ), and each column denotes a region of 5 k . It is clear that the digital ReRAM crossbar (only 500 k and 5M ) can achieve a Table 6.6 Bitwise CNN configuration of MNIST data set Name
Filter/weight
Output size
CONV-1 MP-1 CONV-2 MP-2 FC-1 FC-2
5 × 5 × 1 × 20 2×2 5 × 5 × 20 × 50 2×2 2, 450 × 500 500 × 10
20 × 28 × 28 20 × 14 × 14 50 × 14 × 14 50 × 7 × 7 500 10
30% SET, Vw = 1.0 V SET, Vw = 0.8 V
25%
SET, Vw = 0.4 V RESET, Vw = –0.5 V
Distribution
20%
RESET, Vw = –0.7 V RESET, Vw = –1.0 V
15%
10%
5%
0%
1
2
3 Resistance (MΩ)
4
5
Figure 6.10 ReRAM configuration with voltage amplitude variation following Gaussian distribution
ReRAM-based machine learning 100%
98.3% 98.0% Conventional CNN on CIFAR-10 Bitwise CNN on CIFAR-10 Conventional CNN on MNIST Bitwise CNN on MNIST
Accuracy
96% 92%
(a)
60% 53.1% 40% 20%
84.8%
5%
10% 15% 20% 25% 30% (b) Device variation
99.5%
98.6% 91.4%
87.4%
88% 84% 0%
100%
Accuracy
134
0% 0%
92.4%
58.1%
CIFAR-10 on truncated CNN CIFAR-10 on bitwise CNN MNIST on truncated CNN MNIST on bitwise CNN
5% 10% 15% 20% 25% 30% 35% Bit width
Figure 6.11 (a) Accuracy comparison with analog ReRAM under device variation; (b) accuracy comparison with truncated CNN with approximation
better uniformity than the analog ReRAM crossbar. The accuracy comparison against device variation on ReRAM is shown in Figure 6.11(b). The percentage of device variation defines the maximum resistance offset, and each ReRAM device follows standard deviation. Monte-Carlo is applied in the generation of device variation. In CIFAR-10, one can observe that when the device variation (ReRAM resistance value) is more than 8%, there is a large output current error reported. For example, when the device variation reaches 29%, the digital ReRAM can have an accuracy of 87.4% with only 4% decreased compared to no variations, better than the analog one with an accuracy of 84.8% with 7.6% decreased. In MNIST, the digital ReRAM is always better even when the device variation is larger than 27%. Error under approximation For a precision-bit direct-truncated CNN, we use the conventional approach to train the full-precision (32-bit) CNN first, and then decrease the precision of all the weights in the network. The numerical experiment results of weights with different bit-widths are shown in Figure 6.11(b). Here, the weights in the bitwise CNN is 1 bit, whose accuracy of is not changed. In CIFAR-10, although the accuracy of the full precision (32-bit) can reach 92.4% in the direct-truncated CNN, the bit-width influences the accuracy a lot especially when the bit-width is smaller than 10. For example, when the precision decreases to 6-bit, the accuracy drops down to only about 11.8%. In MNIST, the accuracy of the direct-truncated CNN drops significantly when the bit-width is smaller than 6-bit. The results show that the proposed bitwise CNN can perform much better than the direct-truncated CNN. Scalability study To achieve a better energy efficiency of BNN, we do the scalability study to find out the BNN parameters for both good testing accuracy and energy efficiency. We use a four-layer BNN on MNIST benchmark as a baseline (100% energy efficiency, as shown in Table 6.6), and change the number of output maps of Layer 2 (CONV-2, 50) and hidden nodes of Layer 3 (FC-1, 500), as shown in Figure 6.12(a) and (b). For each energy-efficiency configuration, we do a 20-epoch
The mapping of machine learning algorithms on XIMA 10% 20% 40% 60% 100% 200% 300% 400% 500%
Testing error rate
25% 20% 15% 10% 5%
14%
10% 20% 40% 60% 100% 200% 300% 400% 500%
12% Testing error rate
35% 30%
135
10% 8% 6% 4% 2%
0% 0%
5%
10% Epoch
(a)
15%
20%
0% 0% (b)
5%
10% Epoch
15%
20%
9% 8.3%
8%
Testing error rate
7% 2nd layer output maps
6%
3rd layer hidden nodes
5% 4%
3.4% 2.8%
3% 2%
2.0% 0% (c)
100%
200%
300%
400%
500%
Energy consumption
Figure 6.12 Scalability study of testing error rate under different energy consumption with (a) Layer 2 output maps and (b) Layer 3 hidden nodes; (c) comparison with Layer 2 output maps and Layer 3 hidden nodes
training to make a fair comparison. When the number of hidden nodes or output maps decreases, the energy efficiency is better but it will cause higher testing error rate. To summarize the scalability study, Figure 6.12(c) shows that the hidden nodes of Layer 2 is more sensitive to testing accuracy. As a result, increasing the hidden nodes of Layer 2 is better for higher accuracy while decreasing the hidden nodes of Layer 3 is better for energy efficiency. Performance comparison In this section, 1, 000 images with 32 × 32 resolution in CIFAR-10 are selected to evaluate the performance among all implementations. Parameters including binary weights, and LUT for batch normalization have been configured by the training process. The detailed comparison is shown in Table 6.7 with
136
ReRAM-based machine learning
numerical results including area, system throughput, computation time and energy consumption. Figure 6.13 shows a recognition example of CIFAR-10 data set. In the numerical experiment, every batch comprises ten images and the results are calculated in parallel. The detailed result figure is included in the supplementary file. The black square represents to the class that each image belongs. The overall power comparison among all the implementations is shown in Table 6.7 under the similar accuracy of 91.4% in CIFAR-10. Compared to CPU-based and GPU-based bitwise CNN, the proposed digital ReRAM-based implementation can achieve up to four-magnitude smaller power. Moreover, compared to FPGA-based and CMOS-ASIC-based conventional CNN, the digital ReRAM-based implementation is 4, 155× and 62× smaller power. Table 6.7 Performance comparison among different software and hardware implementations on CIFAR-10 BCNN
Implementations Network category Frequency Power Area (mm2 ) System throughput Frame per second Energy efficiency (GOPS/W) Area efficiency (GOPS/mm2 )
CPU (Matconvnet)
GPU [188]
FPGA [229]
CMOSASIC [230]
Analog ReRAM [146]
Digital passive ReRAM
Bitwise CNN 3.46 GHz 130 W 240 1.48 GOPS 1.2 FPS
Bitwise CNN 1.05 GHz 170 W 601 493 GOPS 400 FPS
CNN
CNN
CNN
100 MHz 18.7 W – 62 GOPS 46 FPS
250 MHz 278 mW 12.25 42 GOPS 35 FPS
– – – -
Bitwise CNN 100 MHz 4.5 mW 0.78 792 GOPS 643 FPS
0.011
2.9
3.32
151
2,000
176,000
0.006
0.82
–
3.43
–
1,015
Ten testing images per batch
– refers to the data not reported.
×128
×256
×512
×512 Recognition results of ten classes
×128
×256
×512
×512
×128
×256
×512
×512
Figure 6.13 Image recognition example of CIFAR-10 data set
The mapping of machine learning algorithms on XIMA
137
In addition, we analyze the detailed power characteristics of the proposed digital ReRAM design in Figure 6.14. First, we analyze the power distribution on convolution, batch normalization, pooling and activation, respectively. Figure 6.14(a) shows that 89.05% of power is consumed by convolution, while 9.17% of power is for batch normalization, and the rest only takes 1.78%. Second, we analyze the power distribution on different bitwise CNN layers in Table 6.2. Results in Figure 6.14(b) show that CONV-6 consumes the most power 25.92%, while CONV-4 and CONV-2 also consume more than 23% of the total power. In general, there is over 98% of power consumed by CONV-2 to CONV-6 layers. For the throughput performance, we use GOPS (giga operations per second) to evaluate all the implementations. The proposed digital ReRAM can achieve 792 GOPS, which is 535× and 1.61× better than CPU-based and GPU-based implementations, respectively. It is also 12.78× and 18.86× better than FPGA-based and CMOS-ASIC. For energy efficiency, the digital ReRAM achieves 176 TOPS/W, threemagnitude better than the CMOS-ASIC-based CNN. In area-efficiency comparison, the digital ReRAM is 296× better than CMOS-ASIC. The digital ReRAM is the best among all the implementations on both throughput and efficiency. The design exploration of CIFAR-10 is shown in Figure 6.15. We change the number of CONV-4 output maps to find out the optimized parameter. The normalized energy consumption is referred to the configuration in Table 6.7. The result shows that the accuracy will not increase when the number is larger than 256. When it is lower than 256, the accuracy drops down even though it has better energy consumption and throughput. As a result, the parameters in Table 6.2 are optimal with specifications in Table 6.7.
6.1.5.4 BCNN-based inference acceleration on 1S1R array Baselines and networks For this experiment, different baselines are implemented for comparison using both CIFAR-10 and MNIST [232] benchmarks, as in Section 6.1.5.3. The detail of each baseline is listed as follows: 1.17%
1.78% 9.17%
Others
25.92%
BN
Conv-6
Conv-3 11.93%
12.38%
89.05%
Conv-2
Conv-5
BinConv
(a)
23.85%
Others
(b)
Conv-4
24.75%
Figure 6.14 Power consumption of bitwise CNN for (a) different operations and (b) different layers
ReRAM-based machine learning 95% 90% 85% 80% 75% 70% 65% 60%
Accuracy Energy consumption Throughput
950
1.2 1.1
900
1.0
850
0.9
800
0.8
Throughput (GOPS)
95% 90% 85% 80% 75% 70% 65% 60%
Normalized energy consumption
Accuracy
138
750
0.7 150
200 250 Output maps of CONV-4 in CIFAR-10
300
Figure 6.15 Performance with different number of CONV-4 output maps in CIFAR-10
Bitwise CNN: We implement bitwise CNN on digital ReRAM crossbar with working frequency of 100 MHz. Comparator on each BL is based on the design of [228]. In the 1S1R array, the drift-type ReRAM in [139] is applied with RLRS = 500 k and RHRS = 5 M . Meanwhile, the diffusive-type ReRAM is taken from [137,138] with Ron = 9 k and Roff = 9 k . For performance comparison, CPU implementation is run in Matconvnet on a computer server with 3.46 GHz core and 64.0 GB RAM. The GPU-based design is referred to [188], and a CMOS-ASIC design in 40-nm technology is also realized. Conventional CNN: FPGA-based [229] and CMOS-ASIC-based [234] implementations are compared. Moreover, an analog ReRAM-crossbar-based realization are simulated.
ReRAM device simulation To illustrate the sneak-path-free effect of the 1S1R implementation, we use SPICE simulation with the same input voltages for Figure 3.14(a) and (c). In the write operation, we set the target Ri, j as high resistance state (HRS) and the sneak-path ReRAMs, Ri, j−1 , Ri+1, j−1 , Ri+1, j are HRS, low resistance state (LRS) and LRS, respectively. By applying VWL,i = 1 and VBL, j = 0 with other nets floating, the sneak-path ReRAM Ri, j−1 in direct connection is changed from 5 to 1.4 M , while the one in the 1S1R implementation remains the original resistance shown in Figure 6.16. In the read operation, we set Ri, j as HRS and the sneak-path ReRAMs are all LRS. According to the results in Figure 6.17, the direct connection implementation will read 3RLRS //RHRS due to the sneak-path effect. Though it takes 500 ns to turn on the selector with read voltage, the target ReRAM can be precisely read out.
Resistance (MΩ) voltage (V)
The mapping of machine learning algorithms on XIMA 1.0 0.8 0.6 0.4 0.2 0.0
139
WL voltage
5 4 3 2 1 0 5 4 3 2 1 0
Target ReRAM in direct connection Sneak-path ReRam in direct connection
Target ReRAM in 1S1R Sneak-path ReRam in 1S1R
0
50
100
150 Time (ns)
200
250
300
Figure 6.16 Sneak-path-free effect in 1S1R implementation for write operation
50
HRS/LRS read ratio in 1S1R
40
Resistance (MΩ) ratio
30 20 10 0 40
Read HRS in direct connection Read HRS in 1S1R Read LRS in 1S1R Read LRS in direct connection
30 20 10 0 0
100
200
300 Time (ns)
400
500
600
Figure 6.17 Sneak-path-free effect in 1S1R implementation for read operation Although the 500 ns read time of the proposed 1S1R is much longer than 10 ns of direct connection, the total read energy will not increase much. As shown in Figure 6.17, because the selector in LRS branch can be turned on faster than the one in RHS branch, the RHS/LRS ratio in the read operation can reach 44 instead of the ideal value 10 (5M/500k). As a result, using selector can achieve a larger difference between LRS and RHS readings.
140
ReRAM-based machine learning
In addition, we show the ReRAM resistance variation with different write voltage Vw in Figure 6.18. We use five sets of ideal Vw from 0.6 to 1.0 V following Gaussian distribution (3σ = 5%Vw ), and 2, 000 data points are simulated to show the distribution of configured ReRAM resistance value in each set. Each column in Figure 6.18 represents a resistance range of 10 k . The results show that the digital ReRAM (LRS with 1.0 V here) has a much better uniformity with input variation compared to analog ReRAM like [150], leading a better programming accuracy.
Performance comparison We select CIFAR-10 as the benchmark to evaluate all the implementations. Table 6.8 summarizes the performance under similar classification accuracy of 91.4%. Power consumption Compared to CPU-based and GPU-based bitwise CNN, the proposed digital ReRAM-based implementation can achieve up to 4× smaller power. Compared to simulation result of CMOS-ASIC bitwise CNN, it has 3.17× smaller power. Moreover, compared to FPGA-based and CMOS-ASIC-based conventional CNN, the digital ReRAM-based implementation is 2, 968× and 44.1× smaller power, respectively. In addition, we analyze the detailed power characteristics of the proposed digital ReRAM design in Figure 6.14. First, we analyze the power distribution on convolution, batch normalization, pooling and activation, respectively. Figure 6.14(a) shows that 89.05% of power is consumed by convolution, while 9.17% of power is for batch normalization, and the rest only takes 1.78%. Second, we analyze the power distribution on different bitwise CNN layers in Table 6.2. Results in Figure 6.14(b) show that CONV-6 consumes the most power 25.92%, while CONV-4 and CONV-2 also 30% Write voltage = 0.6 V Write voltage = 0.7 V Write voltage = 0.8 V Write voltage = 0.9 V Write voltage = 1.0 V
Percentage of ReRAM
25%
20%
15%
10% 5%
0% 0.4
0.6
0.8 1.0 1.2 Configured resistance value (MΩ)
1.4
1.6
Figure 6.18 Programming inaccuracy for different target resistances with voltage amplitude variation following Gaussian distribution
The mapping of machine learning algorithms on XIMA
141
Table 6.8 Performance comparison under different software and hardware implementations
Implementations Network category Frequency Power Area (mm2 ) System throughput Energy efficiency (GOPS/W) Area efficiency (GOPS/mm2 )
CPU (Matconvnet)
GPU [188]
FPGA [229]
CMOSASIC [230]
CMOSASIC simulation
Digital 1S1R ReRAM
Bitwise CNN 3.46 GHz 130 W 240 1.48 GOPS 0.011
Bitwise CNN 1.05 GHz 170 W 601 493 GOPS 2.9
CNN
CNN
100 MHz 18.7 W – 62 GOPS 3.32
250 MHz 278 mW 12.25 42 GOPS 151
Bitwise CNN 30 MHz 20 mW 10 621 GOPS 31,000
Bitwise CNN 100 MHz 6.3 mW 1.02 792 GOPS 126,000
0.006
0.82
–
3.43
62.1
776
– refers to the data not reported.
consume more than 23% of the total power. In general, there is over 98% of power consumed by CONV-2 to CONV-6 layers. Throughput and energy efficiency For the throughput performance, we use GOPS to evaluate all the implementations. The proposed digital ReRAM can achieve 792 GOPS, which is 535×, 1.61× and 1.28× better than CPU-based, GPU-based and CMOS-ASIC-based bitwise CNN, respectively. It is also 12.78× and 18.86× better than FPGA-based and CMOS-ASIC-based CNN. For energy efficiency, the digital ReRAM achieves 126 TOPS/W, three-magnitude better than the CMOS-ASICbased CNN. In area-efficiency comparison, the digital ReRAM is 296× better than CMOS-ASIC-based CNN as well. The digital ReRAM is the best among all the implementations on both throughput and efficiency.
6.2 Machine learning algorithms on 3D XIMA In addition, this chapter presents a detailed mapping of machine learning applications on 3D CMOS-ReRAM architecture, including SLFN-based learning and tensorized neural network (TNN) inference.
6.2.1 On-chip design for SLFN The details of this section are obtained from [235]. This work presents mapping the SLFN-based learning and inference processes according to the algorithm and
142
ReRAM-based machine learning
network in Section 4.2. This section will focus on the mapping of all the matrix– vector multiplication operations on ReRAM crossbar and other operations in CMOS. Multilayer 3D CMOS-ReRAM architecture in Figure 5.13 is used for the acceleration. This section will first discuss the data quantization on the proposed architecture. Then this section illustrates how to map matrix–vector multiplication on ReRAMcrossbar layer. Finally, CMOS-ASIC accelerator is also designed in Layer 3 for remaining operations in pipelined and parallel fashion.
6.2.1.1 Data quantization To implement the whole training algorithm on the proposed 3D multilayer CMOSReRAM accelerator, the precision of the real values requires a careful evaluation. Compared to the software double precision floating point format (64-bit), the values in the training algorithm are truncated into finite precision. By cross-validation from software, a balanced point between hardware resource and testing accuracy can be achieved. A general Nb bit fixed point representation with mb bit integer and nb bit fraction point is shown as follows: Nb −2
y = −bNb −1 2mb +
bl 2l−nb
(6.19)
l=0
In the described accelerator, one needs to first assign a fixed 16 bit-width for normalized input data such as X, Y and D with scale factor mb = 0 and nb = 16. For weights such as A and B in SLFN, one can determine mb by finding the logarithm of dynamic range (i.e., log2 (Max − Min)) and nb is set as 16 − mb . Furthermore, one needs to apply greedy search method based on the testing accuracy to find the optimal bit-width. The bit-width is reduced with cross-validation applied to evaluate the effect. By dynamic tuning the bit-width, our objective is to find the minimum bit-width with acceptable testing accuracy loss.
6.2.1.2 ReRAM layer implementation for digitized matrix–vector multiplication The matrix–vector multiplication is illustrated with the example of output layer (4.7) calculation, where each element in Y is calculated using yij = Lk=1 hik γkj . Such a multiplication can be expressed in terms of binary multiplication [147]. ⎞ E−1 ⎛G−1 L γkj yij = Behik 2e ⎝ Bg 2 g ⎠ k=1
=
e=0
L E−1 G−1 e=0 g=0
k=1
g=0 γ
Behik Bgkj 2e+g
=
E−1 G−1
(6.20) seg 2e+g
e=0 g=0
where seg is the accelerated result from ReRAM crossbar. Bhik is the binary bit of hik with E bit-width and Bγkj is the binary bit of γkj with G bit-width. As mentioned above, bit-width E and G are decided using cross-validation in design to achieve balanced accuracy and hardware usage [236].
The mapping of machine learning algorithms on XIMA
143
Step 1: Parallel digitizing: It requires an L × L ReRAM crossbar. Each inner product is produced ReRAM crossbar as shown in Figure 6.19(a). hi is set as crossbar input and γ is written in ReRAM cells. In the L × L ReRAM crossbar, resistance of ReRAMs in each column is the same, but Vth among columns is different. As a result, the output of each column mainly depends on ladder-like threshold voltages Vth, j . If the inner-product result is s, the output of Step 1 is like (1 · · · 1, 0 · · · 0), where O1,s = 1 and O1,s+1 = 0. Step 2: XOR: It is to identify the index of s with the operation O1,s ⊕ O1,s+1 . Note that O1,s ⊕ O1,s+1 = 1 only when O1, j = 1 and O1, j+1 = 0 from Step 1. The mapping of ReRAM-crossbar input and resistance is shown in Figure 6.19(b), and threshold voltage configuration is Vth, j = V2Rr Rons . Therefore, the index of s is identified by XOR operation. Step 3: Encoding: The third step produces s in binary format as an encoder with the thresholds as Step 2. The input from the second step produces (0 · · · 1, 0 · · · 0) like result where only the sth input is 1. As a result, only the sth row is read out and no current merge occurs in this step. The according binary format binary(s) is an intermediate result and stored in the sth row, as shown in Figure 6.19(c). Encoding step needs an L × n ReRAM crossbar, where n = log2 L is the number of bits in order to represent 1 to L in binary format.
6.2.1.3 CMOS layer implementation for decoding and incremental least-squares In the CMOS layer, decoding and more complex operations for incremental leastsquares solution are designed. Since the output from ReRAM layer is in binary format, decoding is required to obtain the real values. Adder and shifter are designed in Layer 3 with CMOS to complete this process as shown in Figure 6.20.
L
O1,0 = ∑hi,k γ k,j
O2,0 = O1,0 +O1,1
k=1
Intermediate result s for yij, binary(s)=O3 Binary of 0 (00)
O1,0
O2,0
O1,1
O2,1
γ3,j
γ3,j O1,1
O2,2
Binary of 1 (01) 0 or 1 Binary of 2(10)
γ4,j
γ4,j
γ4,j
O1,2
O2,3
Binary of 3(10)
γL,j
γL,j
γL,j
O1,L-1
O2,L
Binary of L
O1,0
O1,1
O1,L-1
hi,1 hi,2 hi,3 hi,4
hi,L
γ1,j
γ1,j
γ1,j
γ2,j
γ2,j
γ2,j
γ3,j
(a)
O2,0
O2,1
(b)
O2,L-1
O3,0
(c)
O3,n-1
O3,n
(d)
Figure 6.19 Detailed mapping for digitalized matrix–vector multiplication on ReRAM crossbar: (a) parallel digitizing; (b) XOR; (c) encoding; (d) legend of mapping
1
0
144
ReRAM-based machine learning
To fully map the incremental machine learning on the proposed 3D multilayer CMOS-ReRAM accelerator, there are following operations needed in CMOS-ASIC including sorting in (4.3), nonlinear mapping in (4.4) and division. Figure 6.21 shows the detailed mapping of the supervised classification on the proposed 3D multilayer CMOS-ReRAM accelerator. The ReRAM logic layer (Layer 2) will work as vector core with parallel process elements for multiplication and summation. The scalar core is implemented to perform scalar operation including division. A sequential divider is implemented with five-stage pipelines to reduce critical path latency. Nonlinear mapping such as Sigmoid function for activation is also implemented in LUT (4.4). As a result, the whole training process including both feature extraction and classifier
Result of ReRAM encoding Intermediate
110010001
s1
Intermediate
n bits
s2
010110101
Adder Shift right
n bits
Control 100101110
+ Product
Final result
Intermediate
s16
yij
Inner-product result
Figure 6.20 Inner-product operation on ReRAM crossbar
LFSR
ReRAM data buffer in Layer 1
Output weight
Shared intermediate results $
Input weight
Input data buffer
TSV I/O memory controller
Vector core Parallel
Scalar core Pipelined
Look-up table
± x ÷ ≥ +
Activation matrix H
CMOS logic in Layer 3
Sigmoid function
x
PE PreH ReRAM logic in Layer 2
Figure 6.21 Detailed pipelined and parallel mapping of CMOS-ASIC accelerator
The mapping of machine learning algorithms on XIMA
145
learning can be mapped to the proposed 3D multilayer CMOS-ReRAM accelerator with small accuracy loss accelerator with small accuracy loss.
6.2.2 On-chip design for TNNs TNNs are another prominent class of machine learning architectures in the present time. In this section, we will map a two-layer TNN-based inference processes according to the algorithm and network in Section 4.4. This section also discusses how to utilize the proposed 3D multilayer CMOS-ReRAM architecture to design the TNN accelerator.
6.2.2.1 Mapping on multilayer architecture Let us first discuss the CMOS layer design, which performs the high-level control of TNN computation. Then a highly parallel ReRAM-based accelerator is introduced with the TNN architecture and dot-product engine.
CMOS layer accelerator To fully map TNN on the 3D multilayer CMOS-ReRAM accelerator, the CMOS logic needs to be designed mainly for logic control and synchronization using top-level state machine. It prepares the input data for computing cores, monitors the states of ReRAM logic computation and determines the computation layer of neural network. Figure 6.22 shows the detailed mapping of the TNN on the proposed 3D multilayer CMOS-ReRAM accelerator. This is a folded architecture by utilizing the sequential operation of each layer on the neural network. The testing data will be collected from ReRAM memory through-silicon via (TSV) and then sent into vector core to perform vector–matrix multiplication through highly parallel processing elements in the ReRAM layer. The ReRAM layer has many distributed ReRAM-crossbar structures to perform multiplication in parallel. Then the computed output from ReRAM will be transferred to scalar score to perform accumulations. The scaler core can perform addition, subtraction and comparisons. Then the output from the scalar core will be sent to the sigmoid function model for activation matrix in a pipelined fashion, which performs the computation of (4.4). The activation matrix H will be used for the next layer computation. As a result, the whole TNN inference process can be mapped to the 3D multilayer CMOS-ReRAM accelerator. In addition, to support TNN on ReRAM computation, a dedicated index LUT is needed. Since the weight matrix is actually folded into a high-dimensional tensor as shown in Figure 4.8, a correct index selection function called bijective function is designed. The bijective function for weight matrix index is also performed by the CMOS layer. Based on the top state diagram, it will choose the correct slice of tensor core G i [i, j] by determining the i, j index. Then the correct ReRAM-crossbar area will be activated to perform vector–matrix multiplication.
ReRAM layer accelerator In the ReRAM layer, this section presents the design of ReRAM layer accelerator for highly parallel computation using single instruction multiple data method to support data parallelism.
146
ReRAM-based machine learning LFSR
ReRAM data buffer in Layer 1
Output weight
Input weight
CMOS logic in Layer 3
Shared intermediate results $ TSV I/O memory controller
Input data buffer
Vector core
± x ÷ ≥
+ Sigmoid function
Activation matrix H
Input data block matrices
Pipelined
Parallel llel
Look-up table
Scalar core
x
PE PreH
ReRAM logic in Layer 2
Input data block matrices
Sub-weight vector configurations (blue crossbars) Sub-input matrix configurations (red crossbars) Vector core -- processing elements Vector core -- processing elements
Vector core -- processing elements Encoded ReRAM result summation Merge
Merge
Merge
H-tree-like distribution network
Encoded ReRAM result summation Encoded ReRAM result summation
Merge
Merge
ReRAM storage Hybrid accelerator
Merge
Tensortrain matrix by vector product
Row voltage controller
Figure 6.22 Data control and synchronization on the layer of CMOS with highly parallel ReRAM-based processing elements
Highly parallel TNN accelerator on the ReRAM layer The TNN accelerator has to be designed to support highly parallel tensor-train-matrix-by-vector multiplication by utilizing the associative principle of matrix product. According to (4.28), X (i) needs to be multiplied by d matrices unlike the general neural network. As a result, if
The mapping of machine learning algorithms on XIMA
147
traditional matrix–vector multiplication in serial is applied, data need to be stored in the ReRAM array for d times, which is time-consuming. Since the size of tensor cores in the TNN is much smaller than the weights in the general neural network, multiple matrix–vector multiplication engines can be placed in the ReRAM logic layer. When then input data are loaded, the index of G i can be known. For instance, one needs to compute X ( j)G 1 [i1 , j1 ]G 2 [i2 , j1 ]G 3 [i3 , j1 ]G i [i4 , j1 ] given d = 4 for the summation in (4.28). G 1 [i1 , j1 ]G 2 [i2 , j1 ] and G 3 [i3 , j1 ]G i [i4 , j1 ] in (4.28) can be precomputed in a parallel fashion before the input data X (i) are loaded. As shown in Figure 6.23, the tensor cores (TC1-6) are stored in the ReRAM logic layer. When the input data X (j) come, the index of each tensor core is loaded by the logic layer controllers first. The controller will write the according data from the tensor cores to ReRAM cells. As a result, the matrix–vector multiplication of G i can
Core index
TC4
TC5
×
G1(i1,j1)G2(i2,j1)
Final matrix multiplication Next layer
TC3
Prestore in ReRAM logic layer
Intermediate result
TC2
Intermediate result
Parallel matrix multiplication
Input data X(j)
High-dimension tensor cores G
TC1
H(i)
Logic layer controller
TC6
×
G3(i3, j1)G4(i4, j1)
G5(i5, j1)G6(i6, j1)
Load input data X( j) from external memory Dot-product engines
(a) TC1 TC2 TC3
× × ×
TC5
× TC6
k=1
Xi,1
×
2,0
Ф1,j
Ф1,j
Ф2,j
Ф2,j
Ф3,j
Ф3,j
Xi,4 Ф4,j
Ф4,j
Ф4,j
Xi,N ФN,j O1,0
ФN,j
ФN,j
O1,1
O1,N-1
Xi,2 Ф2,j Xi,3 Ф3,j
×
TC4
Dot-product engine implementation flow N O1,0 = ∑Xi,kФk,j O3 = Binary(s), if O2,s = 1 O = O +O Ф1,j
1,0
1,0
Binary of 0 (00)
O1,0
O2,0
O1,1
O2,1
Binary 0 or 1 of 1 (01)
O1,1
O2,2
Binary of 2 (10)
O1,2
O2,3
Binary of 3 (10)
O1,N-1
O2,N
Binary of N
O2,0
O2,1
O2,N-1
O3,0
O3,n-1
1 0
O3,n
Data (b)
(c)
Figure 6.23 (a) ReRAM-based TNN accelerator; (b) tree-based parallel partial result multiplication; (c) ReRAM-based dot-product engine
148
ReRAM-based machine learning
be performed in parallel to calculate the intermediate matrices while X (i) is in the loading process. After all the intermediate matrices are obtain, they can be multiplied by X (i) so that the operations in ReRAM logic layer can be efficient. Highly parallel dot-product engine on the ReRAM layer We further develop the digitalized ReRAM based dot-product engine on the ReRAM layer. The tensor-trainmatrix-by-vector operation can be efficiently accelerated by the fast dot-product engine on the ReRAM layer, as shown in Figure 6.23(b). By taking correct index of cores, each operation can be divided into a vector-vector dot product operation. Here, we design the dot-product engine based on [237]. The output matrix is Y , with input matrices X and for better understanding. The overall equation is Y = X with Y ∈ RM ×m , X ∈ RM ×N and ∈ RN ×m . For every element in Y, it follows yij =
N
xik ϕkj
(6.21)
k=1
where x and ϕ are the elements in X and , respectively. The basic idea of implementation is to split the matrix–vector multiplication to multiple inner-product operations of two vectors xi and ϕj . Such multiplication can be expressed in binary multiplication by adopting fixed-point representation of xik and ϕkj : ⎞ E−1 ⎛G−1 N γkj Behik 2e ⎝ Bg 2 g ⎠ yij = k=1
=
e=0
E−1 G−1 e=0 g=0
g=0 N k=1
γ
Behik Bgkj 2e+g
=
E−1 G−1
(6.22) seg 2e+g
e=0 g=0
where seg is the accelerated result from ReRAM crossbar. Bhik is the binary bit of hik with E bit-width and Bγkj is the binary bit of γkj with G bit-width. As mentioned above, bit-width E and G are decided during the algorithm level optimization. The matrix–vector multiplication based on (6.22) can be summarized in four steps on the ReRAM layer. Step 1: Index bijection: Select the correct slice of tensor cores G d [id , jd ] ∈ Rrd ×rd+1 , where a pair of [id , jd ] determines a slice from G d ∈ Rrd ×nd ×rd+1 . In the current example, the X and to represent two selected slices from cores G1 and G2 . Step 2: Parallel digitizing: It requires an N × N ReRAM crossbar. Each inner product is produced ReRAM crossbar as shown in Figure 6.23(c). hi is set as crossbar input and γ is written in ReRAM cells. In the N × N ReRAM crossbar, resistance of ReRAMs in each column are the same, but V th among columns are different. As a result, the output of each column mainly depends on ladder-like threshold voltages Vth, j . If the inner-product result is s, the output of step 1 is like (1 · · · 1, 0 · · · 0), where O1,s = 1 and O1,s+1 = 0. Step 3: XOR: It is to identify the index of s with the operation O1,s ⊕ O1,s+1 . Note that O1,s ⊕ O1,s+1 = 1 only when O1, j = 1 and O1, j+1 = 0 from Step 1. The mapping of ReRAM-crossbar input and resistance is also shown in 6.23(c), and threshold
The mapping of machine learning algorithms on XIMA
149
voltage configuration is Vth, j = V2Rr Rons . Therefore, the index of s is identified by XOR operation. Step 4: Encoding: The third step produces s in binary format as an encoder with the thresholds as Step 2. The input from the second step produces (0 · · · 1, 0 · · · 0) like result where only the sth input is 1. As a result, only the sth row is read out and no current merge occurs in this step. The according binary format binary(s) is an intermediate result and stored in the sth row, as shown 6.23(c). Encoding in Figure step needs an N × n ReRAM crossbar, where n = log2 N is the number of bits in order to represent 1 to N in binary format. By applying these four steps, one can map different tensor cores on ReRAM crossbars to perform the matrix–vector multiplication in parallel. Compared to the stateof-the-art realizations, this approach can perform the matrix–vector multiplication faster and more energy-efficient.
6.2.2.2 Mapping TNN on single-layer architecture Further, this section shows how to map a two-layer TNN-based inference processes according to the algorithm and network in Section 4.4. This section further discusses how to utilize the 3D single-layer CMOS-ReRAM architecture to design the TNN accelerator. Compared to the mapping in Section 6.2.2, a higher parallelism can be achieved with a distributed mapping.
TNN accelerator design According to (4.28), X (i) needs to be multiplied by d matrices unlike the general neural network. In this equation, it is only needed to take a slice of 3D matrix (i.e., a 2-D matrix) from each tensor core to perform the matrix–vector multiplication. Therefore, the computation complexity can be significantly reduced. We show the detailed accelerator design between two hidden layers of TNN in Figure 6.24. Since the parameters of the neural network are compressed, one can store all the weights of TNN in ReRAM crossbar, following Figure 5.12(b). The input data H 1 is processed by several tensor cores, CMOS pooling and activation serially. The detailed design of a tensor core is also shown in Figure 6.24. In each tensor core, one has to store different slices of the 3D matrix into different ReRAM crossbars. Since only one 2D matrix is used at a time, two tensor-core Multiplexers (MUX) are used so that only one matrix is connected to the input voltage as well as the output ADC. The TC selection module controls the input and output MUX according to i and j. For the tensor cores in the middle, the ADC/DAC pair can be removed since one does not need a digital representation for ReRAM-crossbar inputs.
Benefits summary of TNN on 3D CMOS-ReRAM architecture From the algorithm optimization perspective, such tensorization can benefit of implementing large neural networks. First, by performing tensorization, the size of neural network can be compressed. Moreover, the computation load can also be reduced by adopting small tensor ranks. Second, a tensorization of weight matrix can decompose the big matrix into many small tensor-core matrices, which can effectively reduce the
150
ReRAM-based machine learning Hidden layers
× TC2
× TC3
TC4
ReRAM-based matrix–vector multiplications
Hidden layer H2
TC1
Activation
×
×
Pooling
Hidden layer H1
O
ut
pu
t
Multiplication Pooling Activation
CMOS
G2(n2,l2)
TC3 DAC bank
G3(2,1)
Output I3=I2×G3
G 3(1,1)
TC3 output MUX
TC3 input MUX
G2 (2,1)
TC3 DAC bank
G 2(1,1)
n3×l3 TC3 matrices
ADC/DAC pair removable for intermediate TC
TC2 DAC bank
n2×l2 matrices
TC2 output MUX
TC2 input MUX
TC2 DAC bank
TC2
TC3 index i3 selector Output I2=H2×G2
TC2 index i2 selector
G3(n3,l3)
Figure 6.24 Mapping details of highly parallel 3D single-layer CMOS-ReRAM accelerator for TNN configuration time of ReRAM. Lastly, the multiplication of small matrix can be performed in a highly parallel fashion on ReRAM to speed up the large neural network processing time. From the hardware computing architecture perspective, the proposed 3D CMOSReRAM accelerator can greatly improve the energy efficiency and neural network processing speed. The matrix–vector multiplication can be intrinsically implemented by ReRAM crossbar. Compared to CMOS, a multibit tensor-core weight can be represented by a single ReRAM, and the addition can be realized by current merging. Moreover, as all the tensor cores for the neural network weights are stored in the ReRAM devices, better area efficiency can be achieved. The power consumption of the proposed CMOS-ReRAM is much smaller than the CMOS implementation due to nonvolatile property of ReRAM. In the TNN accelerator design shown in Figure 6.24, only a slice of the 3D matrix is active while most of the matrices are inactive. For traditional CMOS realization, it costs huge static power for the inactive tensor-core weights. But in the discussed architecture, the inactive matrices do not consume power so that the energy utilization is high, resulting a better energy efficiency.
The mapping of machine learning algorithms on XIMA
151
6.2.3 Experimental evaluation of machine learning algorithms on 3D CMOS-ReRAM 6.2.3.1 On-chip design for SLFN-based face recognition Experimental settings In the experiment, different baselines are implemented for performance comparisons. The detail of each baseline is listed as follows: Baseline 1: General processor. The general process implementation is based on MATLAB on a computer server with 3.46 GHz core and 64.0 GB RAM. Baseline 2: GPU. The GPU implementation is performed by MATLAB GPU parallel toolbox on the same server. An Nvidia GeForce GTX 970 is used for the acceleration of matrix–vector multiplication operations for learning on neuron network. Baseline 3: 3D-CMOS-ASIC. The 3D-CMOS-ASIC implementation with proposed architecture is done by Verilog with 1 GHz working frequency based on CMOS 65nm low power PDK. Power, area and frequency are evaluated through Synopsys DC compiler (D-2010.03-SP2). TSV model in [238] is included for area and power evaluation. 512 vertical TSVs are assumed between layers to support communication and parallel computations [239]. Proposed 3D-ReRAM-CMOS: The settings of CMOS evaluation and TSV model are the same as Baseline 3. For the ReRAM-crossbar design evaluation, the resistance of ReRAM is set as 1k and 1 M as on-state and off-state resistance respectively according to [206] with working frequency of 200 MHz.
Performance evaluation Scalability analysis To evaluate the proposed 3D multilayer CMOS-ReRAM architecture, we performed the scalability analysis of energy, delay and area on glass UCI data set [184]. In SLFN, the number of hidden nodes may change depending on the accuracy requirement. As a result, the improvement of proposed accelerator with different L from 64 to 256 is evaluated as shown in Figure 6.25. With the increasing L, more computing units are designed in 3D-CMOS-ASIC and ReRAM crossbar to evaluate the performance. When L reaches 256, ReRAM crossbar can achieve 2.1× area-saving and 10.02× energy-saving compared to 3D-CMOS-ASIC. In Figure 6.25(d), energy-delay-product (EDP) of ReRAM crossbar increases faster than 3D-CMOS-ASIC. As the size of ReRAM crossbar is proportional to square of L, the EDP improvement of ReRAM crossbar is less with larger L. However, it still shows great advantage in EDP with 51× better than 3D-CMOS-ASIC when L = 500, which is large number of hidden node for glass benchmark of nine features. Bit-width configuration analysis Table 6.9 shows the testing accuracy under different data sets [184,240], and configurations for machine learning of support vector machine (SVM) and SLFN. It shows that the accuracy of classification is not very sensitive to the ReRAM configuration bits. For example, the accuracy of Iris data set is working with negligible accuracy at 5 ReRAM bit-width. When the ReRAM bit-width increased to 6, it performs the same as 32 bit-width configurations. Similar observation is found in [236] by truncating algorithms with limited precision for better
152
ReRAM-based machine learning
1,000
1
Area (mm2)
CPU CMOS accelerator Hybrid accelerator
10
Delay (s)
0.1 100
0.01 CPU CMOS accelerator Hybrid accelerator
1E-3 1 50
100
(a)
150 200 Hidden nodes L
250
1E-4 50 (b)
Energy (J)
10
CPU CMOS accelerator Hybrid accelerator
1 0.1
EDP (J˙S)
100
0.01 1E-3 50 (c)
100
150 200 Hidden nodes L
100 10 1 0.1 0.01 1E-3 1E-4 1E-5 1E-6
(d)
250
CPU CMOS accelerator Hybrid accelerator
50
250
100 150 200 Hidden nodes L
100
150
200
250
Hidden nodes L
Figure 6.25 Scalability study of hardware performance with different hidden node numbers for (a) area, (b) delay, (c) energy and (d) EDP
energy efficiency. Please note that training data and weight-related parameters are quantized to perform matrix–vector multiplication on ReRAM-crossbar accelerator. Figure 6.26 shows the energy comparisons under different bit-width configurations for CMOS and ReRAM under the same accuracy requirements. An average of 4.5× energy saving can be achieved for the same number of bit-width configurations. The energy consumption is normalized by the CMOS 4 bit-width configuration. Furthermore, we can observe that not always smaller number of bits achieves better energy saving. Fewer number of bit-width may require much larger neuron network to perform required classification accuracy. As a result, its energy consumption increases. Performance analysis Figure 6.27 shows the classification values on image data [240] with an example of five classes. As mentioned before, the index with maximum values (highlighted in red) is selected to indicate the class of test case. A few sample images are selected. Please note that 50, 000 and 10, 000 images are used for training and testing with ten classes. In Table 6.10, performance comparisons among MATLAB, 3D-CMOS-ASIC and 3D multilayer CMOS-ReRAM accelerator are presented, and the acceleration of each procedure based on the formula described is also addressed.
The mapping of machine learning algorithms on XIMA
153
Table 6.9 Testing accuracy of machine learning techniques under different data set and configurations (normalized to all 32 bits) Accuracy (%) Data sets
Size
Features
Glass
214
9
6
Iris
150
4
3
Seeds
210
7
3
Arrhythmia
179
13
3
Letter
20,000
16
7
CIFAR-10
60,000
1,600†
† ‡
Classes
10
Bit-width
SVM‡
SLFN
4 5 6 4 5 6 4 5 6 4 5 6 4 5 6 4 5 6
100.07 93.88 99.82 98.44 100.00 100.00 97.98 99.00 99.00 96.77 99.18 99.24 97.26 98.29 99.55 98.75 99.31 99.31
100.00 99.30 99.30 94.12 94.18 100.00 82.59 91.05 98.51 97.67 98.83 100.00 53.28 89.73 96.13 95.71 97.06 99.33
1,600 features extracted from 60, 000 32 × 32 color images with ten classes. Least-squares SVM is used for comparison.
Normalized energy consumption
1,000 CMOS (4 bits) RRAM (4 bits) CMOS (5 bits) RRAM (5 bits) CMOS (6 bits) RRAM (6 bits)
100
10
1
0.1
0.01 GlassR
Iris
Seeds
Arrhythmia
Letter
Data set
Figure 6.26 Energy-saving comparison under different bit-width of CMOS and ReRAM with the same accuracy requirement
154
ReRAM-based machine learning
Test cases/ Classes
Test case 1 Test case 2 Test case 3
Class 2 dog
Class 1
–3.3363
–0.0037
Class 2
–3.1661
Class 3
–3.5008
0.6737 0.0613
Class 4
–3.4521
–0.0527
1.5498
–3.0085
–0.1861
1.0764
Class 5
Class 1 ship
2.2211 1.2081
Class 3 airplane
1.4670
Class 4 bird Class 5 cat
Figure 6.27 Online machine learning for image recognition on the proposed 3D multilayer CMOS-ReRAM accelerator using benchmark CIFAR-10
Table 6.10 Performance comparison under different software and hardware implementations
Implementation CPU (MATLAB)
3D-CMOSASIC†
3D-CMOSReRAM†
Hardware specifications 3.46 GHz 240 mm2 130 W
1.0 GHz 1.86 mm2 39.41 W
1.0 GHz 1.86 mm2 39.41 W
Computation
Operations
Time
Energy
Feature extraction Input layer L2-norm
Sort Multiplication Multiplication Division Multiplication Multiplication – Sort Multiplication Multiplication Division Multiplication Multiplication – – Sort Multiplication Multiplication Division Multiplication Multiplication – –
1473.2 s 736.6 s 729.79 s 294.34 s 1667.95 s 750.3 s 5651.88 s 216.65 s 97.43 s 96.53 s 43.29 s 220.62 s 99.25 s 773.77 s 7.3× 216.65 s 22.43 s 22.22 s 43.29 s 50.79 s 22.85 s 378.22 s 14.94×
191.52 kJ 95.76 kJ 94.87 kJ 38.26 kJ 216.83 kJ 97.54 kJ 734.78 kJ 78 J 3, 840 J 3, 800 J 15 J 8, 690 J 3, 910 J 20.34 kJ 36.12× 78 J 293 J 291 J 15 J 670 J 300 J 1, 643 J 447.17×
Output layer Overall Feature extraction Input layer L2-norm Output layer Overall Improvement Feature extraction Input layer L2-norm Output layer Overall Improvement
†
6 bit-width configuration is implemented for both CMOS and ReRAM.
The mapping of machine learning algorithms on XIMA
155
Among the three implementations, 3D multilayer CMOS-ReRAM accelerator performs the best in area, energy and speed. Compared to MATLAB implementation, it achieves 14.94× speedup, 447.17× energy-saving and 164.38× area-saving. We also design a 3D-CMOS-ASIC implementation with similar structure as 3D multilayer CMOS-ReRAM accelerator with better performance compared to MATLAB. The proposed 3D multilayer CMOS-ReRAM 3D accelerator is 2.05× speedup, 12.38× energy-saving and 1.28× area-saving compared to 3D CMOS-ASIC. To compare the performance with GPU, we also implemented the same code using MATLAB GPU parallel toolbox. It takes 1163.42s for training benchmark CIFAR-10, which is 4.858× faster than CPU. Comparing to our proposed 3D multilayer CMOS-ReRAM architecture, our work is 3.07× speedup and 162.86× energy saving. Detailed comparisons of each step are not shown due to the limited space of table.
6.2.3.2 Results of TNN-based on-chip design with 3D multilayer architecture Experimental settings In the experiment, we have implemented different baselines for performance comparisons. The detail of each baseline is listed as follows: Baseline 1: General CPU processor. The general process implementation is based on MATLAB with optimized C-program. The computer server is with 6 cores of 3.46 GHz and 64.0 GB RAM. Baseline 2: General GPU processor. The general-purpose GPU implementation is based on the optimized C-program and MATLAB parallel computing toolbox with CUDA-enabled Quadro 5000 GPU [241]. Baseline 3: 3D CMOS-ASIC. The 3D CMOS-ASIC implementation with proposed architecture is done by Verilog with 1 GHz working frequency based on CMOS 65nm low power PDK. Power, area and frequency are evaluated through Synopsys DC compiler (D-2010.03-SP2). TSV area, power and delay are evaluated based on Simulator DESTINY [218] and fine-grained TSV model CACTI-3DD [242]. The buffer size of the top layer is set 128 MB to store tensor cores with 256 bits data width. The TSV area is estimated to be 25.0 μm2 with capacitance of 21 fF. Proposed 3D CMOS-ReRAM: The settings of CMOS evaluation and TSV model are the same as Baseline 2. For the ReRAM-crossbar design evaluation, the resistance of ReRAM is set as 500 k and 5 M as on-state and off-state resistance and 2V SET/RESET voltage according to [139] with working frequency of 200 MHz. The CMOS and ReRAM integration is evaluated based on [64]. To evaluate the proposed architecture, we apply UCI [184] and MNIST [232] data set to analyze the accelerator scalability, model configuration analysis and performance analysis (Table 6.11). The model configuration is performed on MATLAB first using tensor-train toolbox [204] before mapping on the 3D CMOS-ReRAM architecture. To evaluate the model compression, we compare our method with support vector decomposition (SVD)-based node-pruned method [243] and general neural network [53]. The energy consumption and speedup are also evaluated. Note that the code for performance comparisons is based on optimized C-Program and deployed as the mex-file in the MATLAB environment.
156
ReRAM-based machine learning
Table 6.11 Bandwidth improvement under different number of hidden nodes for MNIST data set Hidden node† (L) Memory required (MB) Memory set (MB) Write bandwidth imp. Read bandwidth imp. †
256 1.025 2M 1.14% 5.02%
512 2.55 4M 0.35% 6.07%
1,024 7.10 8M 0.60% 9.34%
2,048 22.20 32M 3.12% 20.65%
4,096 76.41 128M 6.51% 51.53%
Four-layer neural network with three FC layers 784 × L, L × L and L × 10.
Performance evaluation 3D Multilayer CMOS-ReRAM accelerator scalability analysis Since neural network process requires frequent network weights reading, memory read latency optimization configuration is set to generate ReRAM memory architecture. By adopting 3D implementation, simulation results show that memory read and write bandwidth can be significantly improved by 51.53% and 6.51%, respectively, comparing to 2D implementation. For smaller number of hidden nodes, read/write bandwidth is still improved but the bottleneck shifts to the latency of memory logic control. To evaluate the proposed 3D multilayer CMOS-ReRAM architecture, we perform the scalability analysis of energy, delay and area on MNIST data set [232]. This data set is applied to multilayer neural network and the number of hidden nodes may change depending on the accuracy requirement. As a result, the improvement of proposed accelerator with different L from 32 to 2,048 is evaluated as shown in Figure 6.28. With the increasing L, more computing units are designed in 3D CMOS-ASIC and ReRAM crossbar to evaluate the performance. The neural network is defined as a four-layer network with weights 784 × L, L × L and L × 10. For computation delay, GPU, 3D CMOS-ASIC and 3D CMOS-ReRAM are close when L = 2,048 according to Figure 6.28(b). When L reaches 256, 3D CMOS-ReRAM can achieve 7.56× area-saving and 3.21× energy-saving compared to 3D CMOS-ASIC. Although the computational complexity is not linearly related to the number of hidden node numbers, both energy consumption and EDP of ReRAM crossbar increase with the rising number of hidden node. According to Figure 6.28(d), the advantage of the hybrid accelerator becomes smaller when the hidden node increases, but it can still have a 5.49× better EDP compared to the 3D CMOS-ASIC when the hidden node number is 2,048. Model configuration analysis Figure 6.29 shows the testing accuracy and running time comparisons for MNIST data set. It shows a clear trend of accuracy improvement with increasing number of hidden nodes. The running time between TNN and general neural network is almost the same. This is due to the relative large rank r = 50 and computation cost of O(dr 2 n max(N , L)). Such tensor-train-based neural network achieve 4× and 8.18× model compression within 2% accuracy loss under 10,24 and 20,48 number of hidden nodes, respectively. Details on model compression are shown in Table 6.12. From Table 6.12, we can observe that the compression rate is directly
The mapping of machine learning algorithms on XIMA 103
101 100
100 CPU GPU CMOS accelerator Hybrid accelerator
Delay (s)
Area (mm2)
102
10–1
10–1
10–2
CPU GPU CMOS accelerator Hybrid accelerator
10–1 10–1 32
64
10–3
128 256 512 1024 2048 Hidden node
(a)
CPU GPU CMOS accelerator Hybrid accelerator
EDP (J·s)
Energy (J)
101
10–1 10–2 10–3
32
(c)
64
32
64
128 256 512 1024 2048 Hidden node (d)
128 256 512 1024 2048 Hidden node
(b)
102
100
157
104 103 102 101 100 10–1 10–2 10–3 10–4 10–5
CPU GPU CMOS accelerator Hybrid accelerator
32
64
128 256 512 1024 2048 Hidden node
1.4
105%
1.2
90%
1.0 0.8
75%
Around 2% accuracy loss 4x model compression
60%
0.6
45%
TTN test time Gen. testing time
0.4
30%
TTN accuracy 0.2
Gen. accuracy 400
800
1,200
1,600
Testing accuracy
Testing time
Figure 6.28 Scalability study of hardware performance with different hidden node numbers for (a) area, (b) delay, (c) energy and (d) EDP
15% 2,000
Number of hidden nodes
Figure 6.29 Testing time and accuracy comparison between TTN and general neural network (Gen.) with varying number of hidden nodes
158
ReRAM-based machine learning
Table 6.12 Model compression under different number of hidden nodes and tensor ranks on MNIST data set Number of hidden nodes† Compression Rank‡ Compression Accuracy ( %) ‡ †
32 5.50 15 25.19 90.42
64 4.76 20 22.51 90.56
128 3.74 25 20.34 91.41
256 3.23 30 17.59 91.67
512 3.01 35 14.85 93.47
1,024 4.00 40 12.86 93.32
2,048 8.18 45 8.63 93.86
Number of hidden nodes are all fixed to 2,048 with four FC layers. All tensor rank is initialized to 50.
Table 6.13 Testing accuracy of learning techniques under different data set and bit-width configuration
Data sets
32-bit acc. (%) & compr.
4-bit acc. (%) & compr.
5 bit acc. (%) & compr.
6 bit acc. (%) & compr.
Glass Iris Diabetes Adult Leuke MNIST
88.6 96.8 71 78.8 88.9 94.38
89.22 95.29 69.55 75.46 85.57 91.28
83.18 96.8 69.4 78.15 87.38 92.79
88.44 96.8 71.00 78.20 88.50 94.08
1.32 1.12 3.14 1.87 1.82 4.18
10.56 8.96 25.12 14.96 14.56 33.44
8.45 7.17 20.10 11.97 11.65 26.75
7.04 5.97 16.75 9.97 9.71 22.29
connected with the rank r, where the memory storage can be simplified as dnr 2 from d k=1 nk rk−1 rk but not directly linked to the number of hidden nodes. One can also observe that by setting tensor-core rank to 35, 14.85× model compression can be achieved with acceptable accuracy loss. Therefore, initialization of a low-rank core and the SVD split of supercore in MALS algorithm are important steps to reduce the core rank and increase compression rate. Bit-width configuration analysis To implement the whole neural network on the proposed 3D multilayer CMOS-ReRAM accelerator, the precision of real values requires a careful evaluation. Compared to the software double-precision floating point format (64-bit), values are truncated into finite precision. By using the greedy search method, an optimal point for hardware resource (small bit-width) and testing accuracy can be achieved. Our tensor-train-based neural network compression techniques can work with low-precision value techniques to further reduce the data storage. Table 6.13 shows the testing accuracy by adopting different bit-width on UCI data sets [184] and MNIST [232]. It shows that accuracy of classification is not very sensitive to the ReRAM configuration bits for UCI data set. For example, the accuracy of Iris data set is working well with negligible accuracy at 5 ReRAM bit-width. When the ReRAM bit-width increased to 6, it performs the same as 32 bit-width configurations. Please note that the best configuration of quantized model weights varies for different data sets and requires careful evaluation.
The mapping of machine learning algorithms on XIMA
159
Performance analysis In Table 6.14, performance comparisons among C-Programoptimized CPU performance, GPU performance , 3D CMOS-ASIC and 3D multilayer CMOS-ReRAM accelerator are presented for 10, 000 testing images. The acceleration of each layer is also presented for three layers (784 × 2,048, 2,048 × 2,048 and 2,048 × 10). Please note that the dimension of weight matrices are decomposed into [4 4 7 7] and [4 4 8 8] with 6 bit-width and maximum rank 6. The compression rate is 22.29× and 4.18× with and without bit truncation. Here, the compression rate is the ratio of volume of weights before and after compression. Among the four implementations, 3D multilayer CMOS-ReRAM accelerator performs the best in area, energy and speed. Compared to CPU, it achieves 6.37× speedup, 2,612× energy-saving and 233.92× area-saving. For GPU-based implementation, our proposed 3D CMOSReRAM architecture achieves 1.43× speedup and 694.68× energy-saving. We also design a 3D CMOS-ASIC implementation with similar structure as 3D multilayer CMOS-ReRAM accelerator with better performance compared to CPU- and GPUbased implementations. The proposed 3D multilayer CMOS-ReRAM 3D accelerator is 1.283× speedup, 4.276× energy-saving and 9.339× area-saving compared to 3D CMOS-ASIC. The throughput and energy efficiency for these four cases are also summarized in Table 6.14. For energy efficiency, the discussed accelerator can achieve 1,499.83 GOPS/W, which has 4.30× better energy efficiency comparing to 3D the CMOS-ASIC result (347.29 GOPS/W). In comparison to our GPU baseline, it has 694.37× better energy efficiency comparing to NVIDIA Quadro 5000. For a newer
Table 6.14 Performance comparison under different hardware implementations on MNIST data set with 10,000 testing images Hardware Implementation specifications Computations Time CPU [244]
GPU [241]
3D CMOSASIC architecture 3D CMOSReRAM architecture †
3.46 GHz 240 mm2 130W 74.64GOPS 513 MHz 529 mm2 152 W 328.41 GOPS 1.0 GHz 9.582 mm2 1.037 W 367.43 GOPS 100 MHz 1.026 mm2 0.317 W 475.45 GOPS
Layer 1 Layer 2 Layer 3 Overall Layer 1 Layer 2 Layer 3 Overall Layer 1 Layer 2 Layer 3 Overall Layer 1 Layer 2 Layer 3 Overall
0.44 s 0.97 s 0.045 s 1.45 s 0.04 s 0.289 s 2.4 ms 0.33 s 0.032 s 0.26 s 2.4 ms 0.295 s 0.025 s 0.20 s 1.7 ms 0.23 s
Speedup Energy
Energysaving
– – – – 11.06× 3.36× 18.67× 4.40× 13.9× 3.71× 18.4× 4.92× 17.77× 4.80× 25.86× 6.37×
– – – – 9.46× 2.87× 15.96× 3.76× 1, 711.8× 455.4× 2, 262× 604.9× 7, 286× 25.86× 11, 000× 2, 612×
57.23 J 125.74 J 5.82 J 188.8 J 4.05 J 43.78 J 0.3648 J 50.19 J 33.3 mJ 0.276 J 2.6 mJ 0.312 J 7.9 mJ 0.64 J 0.5 mJ 0.072 J
Four-layer neural network with weights 784 × 2,048, 2,048 × 2,048 and 2,048 × 10.
160
ReRAM-based machine learning
GPU device (NVIDIA Tesla K40), which can achieve 1,092 GFLOPS and consume 235 W [241], our proposed accelerator has 347.49× energy-efficiency improvement.
6.2.3.3 TNN-based distributed on-chip design on 3D single-layer architecture Experiment settings In the experiment, we have implemented different baselines for performance comparisons. The detail of each baseline is listed as follows: Baseline 1: General processor (CPU/GPU). The general processor implementation is based on optimized C-program on a computer server with 6 cores of 3.46 GHz and 64.0 GB RAM. A Nvidia GeForce GTX 970 is used as reference for the GPU performance comparison. Baseline 2: 2D CMOS-ASIC. The 2D CMOS-ASIC implementation with proposed architecture is done by Verilog with 1 GHz working frequency based on CMOS 65-nm low-power PDK. Power, area and frequency are evaluated through Synopsys DC compiler (D-2010.03-SP2). Baseline 3: 3D CMOS-ASIC. The 3D CMOS-ASIC implementation is the same as 2D version but features with TSV communication. TSV area, power and delay are evaluated based on Simulator DESTINY [218] and fine-grained TSV model CACTI3DD [242]. The buffer size of the top layer is SRAM based and set 128 MB to store tensor cores. The TSV area is estimated to be 25.0 μm2 with capacitance of 21 fF. TSV model in [238] is included for area and power evaluation; 512 vertical TSVs are assumed between layers to support communication and parallel computations [239]. Proposed 3D CMOS-ReRAM: The settings of CMOS evaluation and TSV model are the same as Baseline 3. For the ReRAM-crossbar design evaluation, the resistance of ReRAM is set as 1 k and 1 M as on-state and off-state resistance and 2V SET/RESET voltage according to [206] with working frequency of 200 MHz. The CMOS and ReRAM integration is evaluated based on [64]. To evaluate the proposed architecture, we apply CIFAR-10 [233] and MNIST [232] data set to analyze the neural network compression and architecture performance analysis. The energy-efficient neural network is designed on MATLAB first using Tensor-train toolbox [204] and tensor network [196]. After determining the network configuration and compression rate, it is then fully optimized using C-program as CPU-based result. Then we apply nonuniform quantization to perform tensor-core quantization before mapping on the 3D CMOS-ReRAM architecture. To evaluate the model compression, we compare our method with SVD-based node-pruned method [243] and general neural network [53]. The energy efficiency and speedup are also evaluated.
Energy-efficient tensor-compressed neural network To achieve high compression rate, we first apply nonuniform quantization for tensorcore weights. As shown in Figure 6.30, the probability density function of tensorcore weights can be modeled as Gaussian distribution. For such known pdf, we can effectively find the optimal level representative values with minimized mean square error. Figure 6.31 shows the trade-off between accuracy, bit-width and compression
The mapping of machine learning algorithms on XIMA 1,200
6 5
Layer 2, tensor cores
800
4
Layer 1, tensor cores
600
3
400
2
200
1
0 –0.5
–0.4
–0.3
–0.2
–0.1 0 0.1 Weight values
0.2
0.3
0.4
Density
1,000
Count
161
0 0.5
Figure 6.30 TNN Layer 1 weight and Layer 2 weight histogram with approximate Gaussian distribution
100%
600
80% Compression rate
400
60%
Accuracy Accuracy loss
300
40% 200
Testing accuracy
Compression rate
500
20%
100
0%
0 0
2
4 6 8 Bit-width of tensor-trai-based weights
10
Figure 6.31 Compression rate and accuracy with increasing bit-width
rate on MNIST data set with three-layer neural network. Table 6.15 shows the clear effect of quantization on the compression rate. In general, bit-width quantization can help improve 3× more compression rate on neural network. To have a fair comparison with [245], we also adopted LeNet-300-100 network [232], which is a two hidden FC layer network (784×300, 300×100, 100×10). Under such configuration, [245] can achieve 40× compression with error rate 1.58% using quantization (6-bit precision), pruning and Huffman coding techniques. By adopting
162
ReRAM-based machine learning
Table 6.15 Model compression under different numbers of hidden nodes and tensor ranks on MNIST data set Rank‡
2
4
6
8
10
12
Compression (×) Q-Compression (×) Error rate (%)
69.73 223.4 4.28
61.70 179.49 4.14
52.53 152.81 4.03
43.54 139.33 4.08
35.71 114.23 3.97
29.30 93.76 3.79
‡
Number of hidden nodes are all fixed to 1,024 with 1 hidden layer.
Table 6.16 Bandwidth improvement under different number of hidden nodes for MNIST data set Hidden node† (L) Memory required (MB) Memory set (MB) Write bandwidth improvement Read bandwidth improvement †
256 1.025 2M 1.14% 5.02%
512 2.55 4 0.35% 6.07%
1,024 7.10 8 0.60% 9.34%
2,048 22.20 32 3.12% 20.65%
4,096 76.41 128 6.51% 51.53%
Four-layer neural network with three FC layers 784×L, L×L and L×10.
9 bit-width configuration on tensorized layer, we can achieve even more compression rate (142×) with maintained accuracy (1.63%). We also apply the proposed TNN to CIFAR-10 data set. CIFAR-10 data set consists of 60, 000 images with size of 32×32×3 under ten different classes. We adopt LeNet-alike neural network [232] and we use a two-layer FC TNN to replace the last two FC layers, which is 512×64 and 64×10. By adopting nonuniform tensorcore bit quantization (8 bit-width), we can achieve 17.6× compression on FC layers and 2× on the whole network, respectively, with the same accuracy loss (2%).
Energy-efficient 3D CMOS-ReRAM architecture Since neural network process requires frequent network weights reading, memory read latency optimization configuration is set to generate ReRAM memory architecture. By adopting 3D implementation, simulation results show that memory read and write bandwidth can be improved by 51.53% and 6.51%, respectively, comparing to 2D implementation as shown in Table 6.16. For smaller number of hidden nodes, read/write bandwidth is still improved but the bottleneck shifts to the latency of memory logic control. In Table 6.17, performance comparisons among general processors, 2D CMOSASIC, 3D CMOS-ASIC and 3D CMOS-ReRAM accelerator are presented, and the acceleration of each layer is also presented for two layers (784 × 10, 24 and 10, 24 × 10). Please note that input weight is decomposed into [4 4 7 7] and [4 4 8 8] with maximum rank 6 and 6 bit-width. Among the four implementations, 3D CMOSReRAM accelerator performs the best in area, energy and speed. Compared to 2D CMOS implementation, it achieves 1.515× speedup, 3.362× energy-saving and
The mapping of machine learning algorithms on XIMA
163
Table 6.17 Performance comparison under different hardware implementations on MNIST data set (10k test images)
Implementations
CPU
GPU
2D CMOSASIC
3D CMOSASIC
3D CMOSReRAM
Frequency Power Area (mm2 ) System throughput Energyefficiency Network compressed Time Energy
3.46 GHz 130 W 240 73.63 GOPS 0.57 GOPS/W Yes
200 MHz 152 W 529 256.32 GOPS 1.69 GOPS/W Yes
1.0 GHz 779 mW 12.69 244.53 GOPS 313.98 GOPS/W Yes
1.0 GHz 649 mW 4.63 286.44 GOPS 441.36 GOPS/W Yes
100 MHz 351 mW 0.61 370.64 GOPS 1, 055.95 W GOPS/W No Yes
1.452 s 188.79 J
0.213 s 32.376 J
0.223 s 0.1738 J
0.191 s 0.1237 J
0.747 s 0.2621 J
0.1473 s 0.0517 J
† 10k test images on three-layer neural network with two FC layers 784×1, 024 and 1, 024×10; 6 bit-width is adopted for tensor-core weights with maximum rank 6.
20.80× area-saving. We also design a 3D CMOS-ASIC implementation with similar structure as 3D CMOS-ReRAM accelerator with better performance compared to 2D CMOS-ASIC. The proposed 3D CMOS-ReRAM accelerator is 1.294× speedup, 2.393× energy-saving and 7.59× area-saving compared to 3D CMOS-ASIC. We also performed the mapping of uncompressed neural network on 3D CMOS-ReRAM architecture. In addition, for CIFAR-10 data set, we estimate 2.79× and 2.23× energy-saving comparing to 2D CMOS-ASIC and 3D CMOS-ASIC architecture, respectively. For energy efficiency, our proposed accelerator can achieve 1,055.95 GOPS/W, which is equivalent to 7.661 TOPS/W for uncompressed neural network. Our proposed TNN accelerator can also achieve 2.39× better energy efficiency comparing to 3D CMOS-ASIC result (441.36 GOPS/W), and 227.24× better energy efficiency comparing to Nvidia Tesla K40, which can achieve 1,092 GFLOPS and consume 235 W.
This page intentionally left blank
Part III
Case studies
This page intentionally left blank
Chapter 7
Large-scale case study: accelerator for ResNet
7.1 Introduction A residual network (ResNet) is a convolutional neural network (CNN)-based network with much deeper hierarchical architecture using the residual connection [246,247], which has become the common network model for deep learning. It is, however, quite challenging to realize the ResNet into terminal hardware with efficient interference [248–251] such as high throughput (TOPs) yet low power (mW). It requires a completed re-examination from both machine learning algorithm as well as underneath hardware. First from machine learning algorithm perspective, to achieve a deep learning at terminal hardware, the network needs to be simplified. For example, a ResNet-50 network contains almost 50 convolution layers, which dominate the network complexity. As most convolution filter now is a small-sized (3 × 3, 5 × 5, etc.) operator. A direct-truncated quantization method has been reported in recent work [252]. Lately, some approaches [253–255] demonstrate their quantization feasibility of using lowbitwidth weights and activations, thus considerably reducing the memory requirement and computation complexity. Low-bitwidth gradients [256] are further generalized to train CNNs. Other works perform the quantization after network training or apply the straight-through estimator on CNNs [257]. Moreover, binarized neural networks (BNNs) are introduced [222,224] and further implemented on a digital resistive random-access memory (ReRAM)-based accelerator. While some of these methods have shown good performances with small-scale network models [258–260], however, none has applied a trained quantization strategy to large-scale networks (e.g., ResNet-50) on large-scale classification tasks (e.g., ImageNet) with certain accuracy. Next, from hardware and device perspective, traditional computing architecture needs significant data migration between memory and processor. It leads to low efficiency for both bandwidth and energy for machine learning, especially the deeper ResNet. The recent emerging nonvolatile memory (NVM) technologies have shown significantly reduced standby power as well as the possibility for in/near-memory computing [225,261–263]. For example, the ReRAM device [57,62,264] in crossbar can be used to store the weights and also to perform the convolution. For this case study, we have developed a quantized large-scale ResNet-50 network using ImageNet [265] benchmark with high accuracy. We further show that
168
ReRAM-based machine learning
the quantized ResNet-50 network can be realized on ReRAM crossbar with significantly improved throughput and energy efficiency. The findings in this chapter can be outlined as follows: 1.
2.
We adopt a trained quantization strategy with all-integer computations of the large-scale deep ResNet, for the first time, to such low-bitwidth subject to an accuracy control. We propose an algorithm-hardware codesign of the complimentary metal oxide semiconductor (CMOS)-ReRAM accelerator with novel digital-to-analog converter (DAC)/analog-to-digital converter (ADC) circuits developed.
Experiment results show that the proposed accelerator can achieve 432× speedup and six-magnitude more energy efficiency than a central processing unit (CPU)based implementation; 1.30× speedup and four-magnitude more energy efficiency compared to a graphics processing unit (GPU)-based implementation; and 15.21× faster and 498× more energy efficiency than a CMOS-application specific integrated circuit (ASIC)-based implementation.
7.2 Deep neural network with quantization The traditional direct implementation of the CNN-based network would require unnecessary software and hardware resources with poor parallelism. Previous work in [256,266] suggests a neural network using quantized constraints during training. In this section, we propose a hardware-friendly algorithm design of the large-scale ResNet model with quantized parallelism for convolution, residual block, batch normalization (BN), max-pooling and activation function. Instead of traditional quantization methods carrying out original 32-bit floating-point arithmetic with additional weight/activation quantization process, we present the all-integer arithmetic with quantized variables and computations only during both training and inference. Due to the remarkable compression on computational complexity and storage size, it becomes feasible to further implement the quantized network on ReRAM crossbar, which will be discussed in Section 7.4.
7.2.1 Basics of ResNet The deep ResNet has achieved outstanding performance in many applications due to its effectiveness. Different from the traditional CNN-based network structure, ResNet utilizes a residual connection to address the degradation problem by training a very deep network. A ResNet is composed of a series of residual blocks, and each residual block contains several stacked convolutional layers. The rectified linear unit (ReLU) layers [267] and the BN layers [190] are also regarded as the component of convolutional layers in a ResNet. Specifically, we adopt a ResNet with 50 layers (ResNet-50) [246] that is substantially deep to extract rich and discriminative features. The specific
Large-scale case study: accelerator for ResNet
169
structure of ResNet-50 is shown in Figure 7.1, including 16 residual blocks and each contains three convolution layers. There is an add-residual operation at the end of each residual block, which can be represented by Ak+2 = F(Ak−1 , Wk , Wk+1 , Wk+2 ) + Ak−1
p In
ut
(7.1)
l ua is d ck e R blo
Ou
conv, 3×3, 128
ut
Image
conv, 1×1, 128 Block#1
tp
CONV1
conv, 1×1, 512
conv, 7×7, 64, /2 Max-pooling, /2 conv, 1×1, 64
conv, 1×1, 128 Block#2
CONV2
conv, 3×3, 128
conv, 1×1, 256
conv, 1×1, 512
3× conv, 1×1, 128
conv, 1×1, 128 Block#3
conv, 3×3, 128
conv, 3×3, 64
CONV3
conv, 3×3, 128 conv, 1×1, 512
conv, 1×1, 512
4× conv, 1×1, 128 Block#4
conv, 1×1, 256
conv, 3×3, 128 conv, 1×1, 512
CONV4
conv, 3×3, 256 conv, 1×1, 1,024 6×
a conv, 1×1, 128 relu F(A,W ) conv, 3×3, 128 relu conv, 1×1, 512 F(A,W )+A + relu
conv, 1×1, 512 CONV5
conv, 3×3, 512 conv, 1×1, 2,048 3× Average-pooling fc 1,000
Figure 7.1 Detailed structure of ResNet-50: 49 convolution layers and 1 FC layer
170
ReRAM-based machine learning
where Ak−1 and Ak+2 denote the input and output feature matrices of the residual block, meanwhile Ak+2 can be considered as the input of the next residual block, and the function F(◦) denotes the residual mapping to be learned.
7.2.2 Quantized convolution and residual block The convolution is the most expensive computation of ResNet and any other CNNbased network. Recently, the low-bitwidth convolution in [188,256] has shown the promising performance, which quantizes the convolution with low-bitwidth parameters to reduce the hardware resource with considerable accuracy. First, we present the low-bitwidth convolution with 4-bit quantized values as {−4, −2, −1, 0, 1, 2, 4} for weight matrices. It should be noted that we quantize weights to the powers of 2. In this way, the quantization algorithm can strike a balance between model compactness and capacity, which has been experimentally verified by testing several realizations from the state-of-the-arts (reported in Section 7.5.3.1). Furthermore, half voltage operating scheme [268] is applied to overcome the sneakpath problem, and the large intervals of weights will help to maintain high robustness under device variation (reported in Section 7.5.3.2). As shown in Figure 7.2(a), assumq ing wk is the full-precision element and wk is the quantized-valued element of weight q matrix Wk in the kth layer, they have the approximation as wk ≈ αwk with a nonnegative scaling factor α. Each weight element in convolution layers can be 4-bit quantized as follows: ⎧ +4, wk > 1.5 ⎪ ⎪ ⎪ ⎪ +2, 1.5 ≥ wk > ⎪ ⎪ ⎪ ⎪ ⎨+1, ≥ wk > 0.5 q |wk | ≤ 0.5 wk = quantize(wk ) = 0, (7.2) ⎪ ⎪ < −0.5 −1, − ≤ w ⎪ k ⎪ ⎪ ⎪ −2, −1.5 ≤ wk < − ⎪ ⎪ ⎩ −4, wk < −1.5
Quantized weight value 4 2 1 –1.5 Δ – Δ –0.5 Δ
0.5 Δ
Δ
1.5 Δ
–1 Original weight value –2 –4
(a)
15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
Quantized feature value × 15
Original feature value × 15 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
(b)
Figure 7.2 4-bit quantization strategy of (a) weight and (b) feature
Large-scale case study: accelerator for ResNet
171
where quantize(◦) represents the proposed 4-bit quantization strategy, and is a nonnegative threshold parameter, which can be calculated as 0.7 |wk | n i=1 n
≈
(7.3)
where n is the number of elements in weight matrix. Also we develop a 4-bit quantization strategy for the features based on [256], which quantizes a real-number feature q element ak−1 to a 4-bit feature element ak−1 of feature matrix Ak−1 in kth layer. This strategy is defined as follows: 1 quantize(ak−1 ) = round[(24 − 1)ak−1 ] (7.4) −1 where the round(◦) function takes the nearest integer (takes the bigger integer if the q decimal part is exactly 0.50). Therefore, the 4-bit quantized feature ak−1 has the value 1 range as 15 × {0, 1, 2, . . . , 15}, which is shown in Figure 7.2(b). Having both 4-bit features and weights, we perform the 4-bit quantized convolution as q
ak−1 =
24
q
sk (x, y, z) =
Dk Hk Vk
q
wk (i, j, l, z)
i=1 j=1 l=1 q
× ak−1 (i + x − 1, j + y − 1, l)
(7.5)
∈ RDk−1 ×Hk−1 ×Vk−1 and ∈ Rdk ×hk ×vk are the 4-bit quantized feature and where q weight element, and sk is the computation result. Here, Dk , Hk and Vk are mode sizes of the features and dk , hk and vk are those of the weight kernels, which can be seen in Figure 7.3. Compared with a real-valued CNN-based network in the full-precision data format, since the weights and features can be stored in 4-bit, both software and hardware resources required for the quantized convolutional layer, called the QConv, can be significantly reduced. Meanwhile, the QConv with 4-bit parameters leads to a higher energy efficiency and parallelism when implemented on the in-memory computing devices. q wk
Inp ut f eat ure s hk
dk
4-bit ResNet
1/1 5x 3 8Dk 11 2 6 4
1 7 5 13 1 2
4-bit quantized features a q
Vk
vk
2 –1 0 2 2 –4 1 4 –1
4-bit quantized weights w q
4-bit quantized BN
Kernels 4-bit quantized ReLUs
Hk
4-bit quantized convolution
In
pu
t
q ak−1
Next 4-bit quantized features
Figure 7.3 Computation flow of deep residual network ResNet-50 with 4-bit quantization
172
ReRAM-based machine learning
For residual block described in (7.1), we perform the proposed optimizations by q straightforwardly replacing the elements in Ak+2 , Ak−1 , Wk , Wk+1 and Wk+2 , as shown in Figure 7.3. Therefore, the add-residual operation at the end of each residual block is also quantized as 4-bit.
7.2.3 Quantized BN BN accelerates the training process and reduces the overall impact of the weight scale [190]. The normalization procedure is required to stabilize the deep ResNet. However, in the original CNN-based network, BN requires many multiplications (calculating the full-precision deviation and dividing by it). In this section, the 4bit quantized BN, called the quantized batch normalization (QBN), is proposed to significantly compress the computation. The quantized computation in QBN can be represented by q |s (x,y,z)| k
qbn
) ak (x, y, z) = 2round(log2 q sk (x, y, z) − μ(x, y, z) × γ (x, y, z) + β(x, y, z) σ 2 (x, y, z)
(7.6)
where μ and σ 2 ∈ RDk ×Hk ×Vk are the expectation and variance over the batch, while γ ∈ R and β ∈ RDk ×Hk ×Vk are learnable parameters that scale the normalized value q |s (x,y,z)| k
) [190], and 2round(log2 are quantized items which shift the normalized value. In 2 the inference stage, μ, σ , γ and β are fixed to normalize the convolution output.
7.2.4 Quantized activation function and pooling The activation function produces a nonlinear decision boundary via nonlinear combinations of the outputs of BN. In the deep ResNet, the ReLU is commonly used which is defined as f (x) = max(0, x) [267]. And the quantized ReLU layers, called the QReLUs, can be optimized as
qbn 0, ak (i, j) ≤ 0 qrelu ak (i, j) = (7.7) qbn round(ak (i, j)), otherwise qbn
qrelu
where ak is the output of the QBN and ak is the quantized activation generated in the kth layer. We can take the QReLUs function for all real-number matrixes using the following example for a 2 × 2 region: −7.63 + 3.11 0 3 =⇒ (7.8) +12.46 − 5.89 12 0 where the 2 × 2 real-value matrix denotes the output of QBN. The QReLUs filter out all the negative values to make the calculation more efficient and more feasible for hardware implementation. Moreover, the pooling operation is optionally applied to select the most significant information from the features, which performs a downsampling across an M × M
Large-scale case study: accelerator for ResNet
173
contiguous region on the feature. There are two kinds of pooling schemes commonly used in CNN-based network: max-pooling and average-pooling. The max-pooling takes the maximum value of the pooling region while the average-pooling takes the mean value of the pooling region. The deep ResNet begins with a max pooling and ends with a global average pooling layer and a 1, 000-way fully connected (FC) layer with a softmax layer. To save runtime memory and storage usage in inference, we can do the fusion as quantize(◦) function commutes with the max-pooling function: quantize(max(x, y)) = max(quantize(x), quantize(y))
(7.9)
and the quantized max-pooling, called the QPooling, is applied with the following equation: qp
qbn
ak (i, j) = max[ak (Mi + k, Mj + l)], (k, l ∈ (0, M )) qbn
where ak
(7.10)
qp
and ak are the features before and after pooling, respectively.
7.2.5 Quantized deep neural network overview The overall working flow of the deep ResNet model with quantization and parallelization is presented in Figure 7.3. After the first layer, the QPooling will primarily work and feed the downsampled features into 4-bit QConv. Then each 4-bit QConv layer conducts a quantized convolution between 4-bit quantized features and weights. Then the feature output of QConv will be further processed by the QBN and QReLUs function. After every three convolution layers, the add-residual operation will be performed with 4-bit quantized parameters at the end of residual block. As all the major parameters and expensive computation are performed with 4-bit quantization, the proposed 4-bit ResNet will get much less storage cost, faster speed as well as higher energy efficiency.
7.2.6 Training strategy The detailed training strategy is presented in Algorithm 7.1. The network is assumed to have a feedforward linear topology, and the computation is performed with 4-bit quantization with high parallelism and efficiency. Note that the quantize(◦) and compute(◦) functions represent the corresponding quantization and computation methods mentioned above, the backward(◦) function specifies how to calculate parameter gradients in the backward propagation [254], and the update(◦) function presents how to update the parameters when their gradients are known [269]. After training, the proposed 4-bit ResNet model can be obtained, which is highly compressed through quantization. Compared to the original ResNet, the quantized ResNet benefits a lot from the large compression and high parallelization. It will reduce the complexity and enlarge the throughput of networks with low-bitwidth parameters, so that the implementation on the in-memory device also becomes more realizable.
174
ReRAM-based machine learning
Algorithm 7.1: Training quantized ResNet with 4-bit weights, features. L is the number of layers and g is the parameter gradient. Require: previous inputs ak−1 and previous weights wk . Ensure: updated weights wk∗ . 1. Computing the forward flow: 1: for k = 1 : L do q 2: wk ← quantize(wk ) q 3: ak−1 ← quantize(ak−1 ) q q 4: sk ← compute(wkb , ak−1 ) qbn q 5: ak ← compute(sk , μ, σ 2 , γ , β) 6: if max-pooling layer then qp qbn 7: ak ← compute(ak ) qrelu qp 8: ak ← compute(ak ) 9: else qrelu qbn 10: ak ← compute(ak ) 11: end if q qrelu 12: ak ← ak 13: end for 2. Computing the backward gradients: 14: for k = L : 1 do q 15: gaq ← backward(gaq , wk ) k−1 k q 16: gwq ← backward(gaq , ak−1 ) k k 17: end for 3. Accumulating the parameters gradients: 18: for k = 1 : L do q 19: wk∗ ← update(wk , gwq ) k 20: end for
7.3 Device for in-memory computing In this section, the ReRAM-crossbar device will be presented which can be used for both storage and computation. And it will be employed to support a CMOS-ReRAM in-memory computing accelerator. The large-scale deep ResNet with quantization which has been discussed in Section 7.2 will be mapped.
7.3.1 ReRAM crossbar ReRAM [62] is an in-memory computing device facing the future. The different resistance states of ReRAM can represent different parameters while computing. Besides working as storage memory, a ReRAM crossbar can be applied to perform logic operations [263], especially the dot-product-based convolutions. Previous works in [258,259] usually choose the two-terminal ReRAM device which only has two nonvolatile states: high-resistance state (HRS) and low-resistance
Large-scale case study: accelerator for ResNet
175
state (LRS) to build an in-memory architecture, as shown in Figure 7.4(a). However, when the applied network becomes deeper and more complex, the two-terminal computing results in limitation on the performance of ReRAM crossbar while sacrifices the accuracy of network. For the implementation of proposed 4-bit ResNet-50, we apply the ReRAM model in [270], which has multibit resistance states. The states of proposed ReRAM device are determined by the distribution of set voltage Vset . In one ReRAM crossbar, given the input probing voltage VW on each writeline (WL), the current IB on each bitline (BL) becomes the natural multiplication– accumulation of current through each ReRAM device. Therefore, the ReRAMcrossbar array can intrinsically perform the matrix–vector multiplication. Given an input voltage vector VW ∈ RN ×1 , the output vector IB ∈ RN ×1 can be expressed as ⎡ ⎤ ⎡ ⎤⎡ ⎤ IB,1 c1,1 · · · c1,M VW ,1 ⎢ .. ⎥ ⎢ .. .. ⎥ ⎢ .. ⎥ .. (7.11) ⎣ . ⎦=⎣ . . . ⎦⎣ . ⎦ IB,M
cN ,1
···
cN ,M
VW ,N
where ci, j is configurable conductance of the ReRAM resistance Ri, j , which can represent real value of weights. Compared to the traditional CMOS implementation, the ReRAM device can achieve a higher level of parallelism and consumes less power including standby power. Different from the traditional non-quantized two-terminal ReRAM crossbar, the proposed 4-bit quantized ReRAM crossbar is shown in Figure 7.4(b), and the 4-bit resistance is configured for each ReRAM device. Moreover, the novel designed 4bit DACs are applied on WL to transfer the digital input, and the 4-bit ADCs are applied on BL to identify the analog output. Therefore, the sensed analog output can be encoded in quantized digital signal to produce the multiplied result [237]. Compared to the two-terminal ReRAM crossbar, the 4-bit ReRAM crossbar has a better robustness on the IR-drop in the large-size implementation, so that the wire resistance affects little on IR-drop. In addition, since only 4-bit inputs are configured
V1
DAC LRS
HRS
LRS
HRS
V2 LRS
HRS
HRS
HRS
V3
RS10
RS1
RS12
RS1
RS3
RS13
RS15
RS9
RS2
RS11
DAC HRS
LRS
LRS
LRS
V4
RS6 DAC
HRS
(a)
RS3 DAC
HRS
HRS
HRS
4-bit RS8 RS12 RS14 RS2 DAC 4-bit ADC ADC (b) ADC ADC
Figure 7.4 (a) Traditional two-terminal ReRAM crossbar; (b) 4-bit ReRAM crossbar with 4-bit DAC/ADC circuits
ADC
176
ReRAM-based machine learning
in quantized ReRAM crossbar, it has a better programming accuracy and can provide a better strategy for mapping deeper and large-scale networks.
7.3.2 Customized DAC and ADC circuits To accommodate the multibit network on the ReRAM crossbar, we customize the designs of the 4-bit DAC and ADC, respectively, as shown in Figure 7.5(a) and (b). To save area, the DAC is implemented with the resistance to resistance (R2R) structure with a compact Opamp buffer. Compared with the traditional resistor ladder, the R2R uses uniform resistor values that favor better matching. The Opamp buffer uses feedforward structure to minimize current consumption and achieves fast settling. Biased by a gm /Id around 15, the current consumption of the whole Opamp is less than 13 μA and achieves a 100 MHz unity gain bandwidth. Meanwhile, an ADC scheme similar to that in [271] is employed to reduce the switching power, which uses asynchronous SAR to minimize the area and energy.
7.3.3 In-memory computing architecture Based on the developed 4-bit ReRAM-crossbar device and novel DAC/ADC circuits, we present the CMOS-ReRAM in-memory computing accelerator with both memory storage and computation implemented as shown in Figure 7.6. In this architecture, data–logic pairs are located in a distributed fashion, where data transmission among data block, logic block and external scheduler are maintained by a control bus. Therefore, logic block can read data locally and write back to the same data block after the logic computation is done [237]. As a result, the huge communication load between memory and general processor can be relieved because most of the data transmission is done inside data–logic pairs. Based on this in-memory computing architecture using the 4-bit CMOS-ReRAM accelerator, we will further introduce the strategy to map the proposed 4-bit quantized ResNet-50.
X3 X2 X1 X0 (a)
1
1
1
R3 V3
+
–
R2 +
V2
Gm3
Vref–n Vcm Vref_p
Gm1
–
R1
Vin_p
Rf
V1 V0
Gm2
–
R0 1
+
Vin_n
Opamp
+ –
Vout
C9 C9
C8 C8
C5 C5
C5r
C1 C0
C5r
C1
C0
Opamp
SAR logic
+ –
(b)
Vref_p Vcm Vref_n
Figure 7.5 Simplified 4-bit (a) DAC and (b) ADC circuits
E/L
Decoder
Large-scale case study: accelerator for ResNet
Driver
In-memory computing
Driver
ReRAM crossbar1 ReRAM crossbar2
Driver
Driver
+ Add residual
Controller
Compiler
177
Residual block
ReRAM crossbar3
In-memory computing Logic block
Data block
Figure 7.6 In-memory accelerator architecture based on 4-bit ReRAM crossbar for 4-bit quantized ResNet-50
7.4 Quantized ResNet on ReRAM crossbar In this section, we present the detailed strategy to map the proposed large-scale ResNet with 4-bit quantization on the 4-bit CMOS-ReRAM, including quantized convolution and add-residual operations.
7.4.1 Mapping strategy In order to avoid mapping negative weights on ReRAM-crossbar device, the 4-bit q weights shown in Figure 7.2(a) are departed into positive weights wkp and negative q weights wkn . Therefore, the quantized convolution can be presented by two separated multiplication operations, which are shown as q
sk =
N N q q q q q q wk · ak−1 = (wkp · ak−1 + wkn · ak−1 ) i=1
i=1 q
ak−1
(7.12)
1 q ∈ {0, 1, 2, . . . , 15}, wk ∈ {0, ±1, ±2, ±4} 15 q
to get the expected calculation result (keep weights positive), and the values of wkp q and wkn will satisfy the following conditions: q q wk , wk ≥ 0 q wkp = (7.13) q 0, wk < 0 q 0, wk ≥ 0 q wkn = (7.14) q q −wk , wk ≤ 0
178
ReRAM-based machine learning Table 7.1 Network model of quantized ResNet-50 Layer/residual block
Output size
Parameters (kernels, channels)
CONV1
112 × 112
CONV2
56 × 56
CONV3
28 × 28
CONV4
14 × 14
CONV5
56 × 56
7 × 7, 64 3 × 3 max pool ⎡ ⎤ 1 × 1, 64 ⎣3 × 3, 64 ⎦ × 3 1 × 1, 256 ⎡ ⎤ 1 × 1, 128 ⎣3 × 3, 128⎦ × 4 1 × 1, 512 ⎡ ⎤ 1 × 1, 256 ⎣3 × 3, 256 ⎦ × 6 1 × 1, 1024 ⎤ ⎡ 1 × 1, 512 ⎣3 × 3, 512 ⎦ × 3 1 × 1, 2048
FC
1×1
7 × 7 average pool softmax output
Such all-positive trick makes it more feasible to implement a large-scale network on ReRAM crossbar. Specifically, the architecture of quantized ResNet-50 is shown in Table 7.1, which will be further mapped on the previously discussed 4-bit ReRAM crossbar.
7.4.2 Overall architecture The whole mapping of residual block of proposed 4-bit ResNet-50 is shown in Figure 7.7. Since there are positive and negative lines for WL, the 4-bit ResNet-50 requires a 2N × N ReRAM crossbar, where N is the number of elements in the feature matrices. The 4-bit DAC and the 4-bit ADC are applied to take input and output features. All columns are configured with the same elements that correspond to the quantized q weight columns wk of the neural network, and the WL voltages are determined by the q analog input ak−1 transferred by the 4-bit DAC. Meanwhile, the output from every q BL will be transferred to the digital signal by a 4-bit ADC. The sk represents the convolution result, which can be seen in (7.5). q As described in (7.5), each quantized weight vector wk performs quantized convolution with several input features. As a result, each logic block in Figure 7.6 stores a q quantized vector wk , while the control bus transmits the input feature sequentially. In this case, quantized convolution can be performed in parallel in separated logic blocks. Although these operations are implemented on different ReRAM-crossbar arrays, the input/output formats of them are compatible so that one can directly connect them as shown in the logic block in Figure 7.6. As discussed in the in-memory computing
Add residual ak-1(1) V–
ak-1(3) ak-1(4)
+ Add residual
ak-1(5)
4-bit ReRAM crossbar
ak-1(N)
ak-1(1) wkp(N)
wkp(N)
wkp(N)
ak-1(2) wkn(N)
wkn(N)
wkn(N)
V+
DAC
V–
ak-1(2)
wkp(N-1) wkp(N-1)
wkn(N-1)
4-bit feature-1
ak-1(N) 4-bit DAC
4-bit ReRAM wkn(1) 4-bit ADC
wkn(N-1)
wkp(N-1)
DAC
DAC
wkn(N-1)
wkn(1)
wkn(1)
ADC
DAC
DAC
QBN, QReLUs
DAC
ak-1(2)
QBN, QReLUs
ak-1(1)
V+
ADC
4-bit result-1 Sk(1) Sk(2)
Sk(3)
ADC Sk(N)
ADC
ADC
4-bit result-2 Sk+1(1) Sk+1(2) Sk+1(3)
ADC Sk+1(N)
ADC
ADC
4-bit result-3 Sk+2(1) Sk+2(2) Sk+2(3)
To next residual block
Figure 7.7
Mapping strategy for a 4-bit quantized ResNet-50 block on 4-bit ReRAM crossbar
ADC Sk+2(N)
180
ReRAM-based machine learning
architecture, one ResNet-50 residual block contains three convolution layers, and the add-residual operation is applied at the end of residual block to straightforwardly add the input of Layer 1 and the output of Layer 3. As a summary, the 4-bit quantized convolution, BN, ReLUs and add-residual operations in Figure 7.1 can be mapped onto the proposed 4-bit CMOS-ReRAM accelerator and the max-pooling operations are performed in the additional processor. The pipeline design is based on Table 7.1. We observe that CONV2, CONV3, CONV4 and CONV5 are the stages which require the most ReRAM cells and computation time. As a result, more logic blocks are assigned to these steps to relieve the critical path. Since the output features from Figure 7.7 are 4-bit quantized, the area overhead and energy consumption of the data storage can be significantly reduced. In addition, as the layers in Table 7.1 are implemented with different data–logic pairs, the volume of data transmitted is also decreased.
7.5 Experiment result 7.5.1 Experiment settings In the experiment, we have implemented different baselines for accuracy and performance comparison. Detailed setting of each baseline is listed as follows: 1.
2. 3. 4. 5.
Proposed 4-bit CMOS-ReRAM: We use the ReRAM model of [272] with uniformly separated resistance values in [1 k, 5 k] and a working frequency of 100 MHz. CPU: The general processor implementation is using Matconvnet [227] on a computer server with 6 cores of 3.46 GHz and 64.0 GB RAM. GPU: The general-purpose GPU implementation is based on a CUDA-enabled video card [188] with 1.05 GHz. State-of-the-art ReRAM-based architectures: Two existing ReRAM-based architectures [146,222]. Others: The field programmable gate array (FPGA) [273], CMOS-ASIC [230] and micron automata processor [274] implementations of the conventional CNN networks.
To present the power of proposed architecture, we apply the large-scale ImageNet [265] data set to analyze the accuracy and performance of proposed 4-bit hardware accelerator. First, the model takes a 224 × 224 × 3 RGB image as the input of the first convolution layer with 64 sets of quantized filters. After that, QPooling is applied to prepare data for the following layers. Next, the QConv layer of the first residual block performs the quantized convolution between input features and weights followed by the QPooling. Then the QBN is employed after the QConv to accelerate the training. Meanwhile, we adopt the QReLUs to each QConv layer to stabilize the process. For every layer stack, three cascaded convolution layers make up a residual block with an add-residual operation. In the end, the output of the last FC layer is fed into a softmax layer to generate a probabilistic distribution of 1, 000 classes.
Large-scale case study: accelerator for ResNet
181
7.5.2 Device simulations According to the sizes of feature matrix in ResNet-50, we employ a 128 × 64 ReRAM crossbar where the number of input features is set as 128 and the number of output channels is 64. We adopted the read/write latency and read/write energy from [275] as 29.31 ns/50.88 ns per spike, 1.08 pJ/3.91 nJ per spike. The final layouts of the proposed 4-bit DAC and ADC circuits are shown in Figure 7.8. We further report simulation results of the proposed CMOS-ReRAM accelerator. In Figure 7.9(a), 4bit digital waveforms are converted into analog signals by the 4-bit DAC. Then in Figure 7.9(b), output waveforms are obtained by the 4-bit ADC for various ReRAM q resistive states representing all possible wkp . A high accuracy can be observed in the DAC/ADC circuits, and the whole CMOS-ReRAM accelerator permits a high performance and efficiency, as shown in Table 7.4.
(a)
(b)
Figure 7.8 Layouts of the proposed 4-bit (a) DAC and (b) ADC
1.2 Rr = 5 kΩ (weight = 4) Rr = 3 kΩ (weight = 2) Rr = 2 kΩ (weight = 1) Rr = 1 kΩ (weight = 0)
400 1 0.8
Voltage (V)
Current (μA)
200
0
0.6 0.4
–200 0.2 DAC output without Opamp DAC output with Opamp
–400 0
(a)
50
100
Time (ns)
150
0 200
0
(b)
100
200
300
400
500
Time (ns)
Figure 7.9 (a) Waveforms of the 4-bit DAC inputs; (b) 4-bit ADC outputs with different ReRAM resistances
600
182
ReRAM-based machine learning
7.5.3 Accuracy analysis First, we report both the top-1 and top-5 accuracy comparison between the proposed 4-bit ResNet-50 and the full-precision ResNet-50 on ImageNet. It should be noted that in case of top-1 rate, we check if the top class having the highest probability is the same as the target label, and in case of top-5 rate, we check if the target label is one of the top five predictions with the highest probabilities. And then, the accuracy comparison under device variation is presented. At last, we report the accuracy comparison between the proposed quantized ResNet-50 and the direct-truncated ResNet-50 under bitwidth approximation.
7.5.3.1 Peak accuracy comparison Figure 7.10 shows the evolution of top-1 and top-5 accuracy vs. epoch curves of the 4-bit ResNet-50 compared with the full-precision ResNet-50. It can be seen that quantizing the ResNet-50 to be 4-bit does not cause the training curve to be significantly different from the full-precision one. Only 3.3% prediction accuracy decreases for top-1 while only 2.5% accuracy losses for top-5 compared with the full-precision ResNet-50. Moreover, we report the top-5 accuracy comparison among different state-of-the-art quantization approaches in Table 7.2. It is observed that the proposed 4-bit trained quantization outperforms other realizations, which is 3.5% higher than the 4-bit fine-grained quantization [276] and 1.5% higher than the 5-bit incremental network quantization [277]. Therefore, we can conclude that the ResNet-50 with the proposed 4-bit trained quantization obtains a remarkable balance between large compression and high accuracy.
7.5.3.2 Accuracy under device variation In the previous discussion, the 4-bit ReRAM crossbar has better programming accuracy than the traditional ReRAM crossbar so that it achieves a better uniformity. For
0.8
1
71.4%
90.6%
0.7
0.9 68.1%
0.8
0.5
Accuracy
Accuracy
0.6
88.1%
0.4 0.3
0.7 0.6 0.5 0.4
Full-precision ResNet-50 4-bit quantized ResNet-50
0.2 0.1 0
(a)
Full-precision ResNet-50 4-bit quantized ResNet-50
0.3 0.2
0
10
20
30
Epoch
40
50
0.1
60
(b)
0
10
20
30
40
Epoch
Figure 7.10 (a) Top-1 validation accuracy curves of quantized ResNet-50. (b) Top-5 validation accuracy curves of quantized ResNet-50
50
60
Large-scale case study: accelerator for ResNet
183
Table 7.2 The top-5 accuracy comparison between the proposed network and state-of-the-art approaches on the ImageNet data set Original ResNet-50 (32-bit) [246] Truncation-based quantization (4-bit) [252] Binary weight quantization (1-bit) [254] Ternary weight quantization (2-bit) [278] Fine-grained quantization (4-bit) [276] Incremental quantization (5-bit) [277] Discussed trained quantization (4-bit)
90.6% 12.3% 78.7% 81.2% 84.6% 86.8% 88.1%
0.92 90.6%
Accuracy
0.9 88.1% 0.88 86.5% 0.86
0.84
Conventional ResNet-50 on GPUs Conventional ResNet-50 on ReRAM 4-bit quantized ResNet-50 on GPUs 4-bit quantized ResNet-50 on ReRAM
0
0.1
0.2
84.7% 0.3
Device variation
Figure 7.11 Accuracy validation of quantized ResNet-50 under ReRAM-crossbar device variation
further evaluation, we build a device variation model by using the Monte-Carlo simulation technique under different writing voltages Vw where the voltage amplitude meets the Gaussian distribution (σ = 1%Vw ). Based on the variation model, we systematically set the parameter variation during the network training and can achieve better accuracy under device variation. The accuracy comparison against device variation on ReRAM is shown in Figure 7.11. It can be observed that when the device variation of ReRAM increases, there are output current errors. For example, when the device variation reaches 29%, the 4-bit ReRAM crossbar has an accuracy of 86.5% with only 1.6% decreases compared with GPUs, better than the non-quantized one with an accuracy of 84.7% with 5.9% decreased. The proposed CMOS-ReRAM delivers higher robustness when the device variation is larger than 20%.
184
ReRAM-based machine learning
7.5.3.3 Accuracy under approximation To evaluate the accuracy under bitwidth approximation, we choose the precision-bit direct-truncated method for comparison. For the proposed trained quantized ResNet50 and direct-truncated ResNet-50, we first train the full-precision (32-bit) network, and then decrease the precision of all the weights and features in the network. The numerical results with different bitwidths are shown in Figure 7.12. In our model, the weights and features in the quantized ResNet-50 are 4-bit, whose accuracy has no significant loss compared with the full-precision ResNet-50. Although the accuracy of the 32-bit model reaches 90.6% in the direct-truncated ResNet-50, the bitwidth influences the accuracy a lot especially when the bitwidth is smaller than 10. For example, when the precision decreases to 6-bit, the accuracy drops down to only about 11.3%. The results show that the proposed quantized ResNet-50 can perform much better than the direct-truncated one. Moreover, we further report the evaluation on accuracy and storage compression among different bitwidths. As shown in Table 7.3, the 4-bit ResNet-50 causes only
90.6%
0.9
Accuracy
0.7 58.1%
0.5
0.3
0.1
Truncated quantization Our trained quantization 5
10
15 20 Bitwidth
25
30
Figure 7.12 Accuracy comparison with direct truncation-based quantization based on ResNet-50 Table 7.3 Accuracy validation of quantized ResNet-50 under bitwidth approximation Bitwidth
Accuracy
Storage compression
2-bit 3-bit 4-bit 6-bit 8-bit 16-bit Full precision
32.9% 71.5% 88.1% 88.7% 89.8% 90.5% 90.6%
15.86 10.85 8.02 5.76 3.94 1.98 N/A
Large-scale case study: accelerator for ResNet
185
0.025 decrease on accuracy with 8.02 storage compression. It can be seen that accuracy decreases a lot when the bitwidth is smaller than 4, but not much improvement when bitwidth is larger than 4. There are some recent works to perform quantization on 8-bit and binary quantization [188]. However, 8-bit quantization results in higher accuracy still have significant hardware cost, while binary quantization cannot achieve a reasonable accuracy on a large-scale neural network. Therefore, we conclude that the 4-bit quantization can have good balance between accuracy and storage compression on the large-scale ResNet-50 network. It is the reason why 4-bit is determined to be the best quantization bitwidth for the large-scale ReRAM implementation.
7.5.4 Performance analysis In the experiments, parameters with quantization have been configured by the training process, meanwhile every batch of inputs and results are calculated in parallel. In Table 7.4, performance comparison among CPU, GPU and other baselines is presented, which is evaluated on 100 testing images with 224 × 224 resolution.
7.5.4.1 Energy and area Using the ReRAM model in [270], the current consumption of each ReRAM cell is less than 1 μA, where the voltage is set to be 5 mV and the area consumption of each ReRAM cell is less than 4 × 4 μm2 . Moreover, the current consumption of the proposed novel DAC/ADC is less than 13 μA and the total area of DAC/ADC is less than 26 × 27 μm2 . Having both settings of ReRAM and DAC/ADC models, the numerical results of energy consumption can be simulated as follows: 5 mV × 0.001 mA × 128 + 5 mV × 0.0122 mA × (128 + 1) = 8.509 mW. It should be noted that we load power on a single BL for each convolution operation, the value 128 denotes the amount of ReRAM cells and (128 + 1) denotes the number of DACs/ADCs in one BL. The energy distribution is shown in Figure 7.13(a) and the convolution is the major operation that dominates 87.4% of the total energy. Also, we calculate the area consumption by 4 μm × 4 μm × 128 × 64 + 26 μm × 27 μm × (128 + 64) = 265, 856 μm2 . Based on the area model in [279], we present the area overhead of the proposed CMOS-ReRAM architecture in Figure 7.13(b). The area overhead of DAC/ADC is the largest part which occupies 32%, and the total area consumption is simulated as 830, 800 μm2 . The energy and area consumption comparison among all the implementations are shown in Table 7.4. It can be seen that the implementation on the proposed 4-bit CMOS-ReRAM accelerator can achieve at least four-magnitude smaller energy consumption compared to GPU and CPU.
7.5.4.2 Throughput and efficiency For the throughput performance, we use GOPS (giga operations per second) to evaluate all the implementations as shown in Tables 7.4 and 7.5. In the throughput comparison, the proposed 4-bit ReRAM crossbar is 432× faster than CPU-based implementation, 10.31× faster than FPGA-based implementation and 15.21× faster than CMOS-ASIC-based implementation. As for energy efficiency, it can be four-magnitude better than GPU-based implementation and 498× better than CMOSASIC-based implementation. And in area-efficiency comparison, it is 224× better
Table 7.4 Performance comparison under different hardware implementations Implementation
CPU [227]
GPU [188]
FPGA [273]
CMOS-ASIC [230]
Network Frequency Area Average power System throughput Energy efficiency Area efficiency
4-bit ResNet-50 3.46 GHz 240 mm2 130 W 1.48 GOPS 0.011GOP/J 0.006 GOP/s/mm2
4-bit ResNet-50 1.05 GHz 601 mm2 170 W 49 GOPS 2.9 GOP/J 0.82 GOP/s/mm2
ResNet-50 240 MHz N/A 25 W 560 GOPS 22.4 GOP/J N/A
Small-scale CNN 250 MHz 12.25 mm2 278 mW 42 GOPS 151 GOP/J 3.43 GOP/s/mm2
Micron automata processor [274] Small-scale CNN 300 MHz N/A 0.5 W 251 GOPS 502 GOP/J N/A
4-bit CMOS-ReRAM 4-bit ResNet-50 100 MHz 0.83 mm2 8.5 mW 639 GOPS 75.2 TOP/J 770 GOP/s/mm2
Large-scale case study: accelerator for ResNet
187
DAC and ADC WDD Control Preprocess Buffer Others Final output
32% Others 12.6%
19% 16% 13% 12% Convolution 87.4%
7% δ . To reduce the quantization error, it is an intrinsic idea to increase the level of quantization. Consider a modified problem formulation ⎧ (kT )ij ≥ 1/2 ⎪ ⎨1, ˜ ij = 0, −1/2 ≤ (kT )ij < 1/2 (8.15) ⎪ ⎩ −1, (kT )ij < −1/2 with each element of the matrix normalized within the interval of [−1, 1]. It is important to keep matrix Boolean so that it can be mapped to ReRAM crossbar struc˜ can be split into two Boolean matrices ture efficiently; thus, it requires the matrix ˜ = 1 ( ˆ1+ ˆ 2 ), where ˜ ∈ {−1, 0, 1} and ˆ 1, ˆ 2 ∈ {−1, 1}. With Boolean quan 2 tization, only one projection ReRAM crossbar is needed. Two ReRAM crossbars are needed for the three-level quantization case, as a result of trade-off between error and hardware complexity.
8.4.3 Overall optimization algorithm The heuristic optimization process is summarized in Algorithm 8.1. Given some ˆ the inner loop of Algorithm 8.1 tries to find the local close-optimal initial guess of , ˆ through iterations. Within each iteration, (8.6) and (8.12) solution by improving are solved by singular vector decomposition and quantization as concluded in (8.10) ˆ stops improving and and (8.15), respectively. The iterations terminate when the converges. As both integer constraint and orthogonal constraint are nonconvex, the local optimum in most cases is not optimal globally. In other words, the solution strongly depends on the initial guess that leads to the local close optimum. Therefore, the outer loop of Algorithm 8.1 increases the search width by generating numerous initial guesses that scatter within orthogonal matrices space. For each initial guess, it will
200
ReRAM-based machine learning
Algorithm 8.1: IH Boolean sampling matrix optimization algorithm Input: Real-valued embedding matrix , search width and quantization level ˆ opt Output: Optimized Boolean embedding matrix ˆ initialize opt ← random m × n Bernoulli matrix 1: while not reach search width limit do 2: seed ← random m × m matrix 3: U, S, V ← SVD of seed 4: T←U 5: while not converged do ˆ ← quantization of T 6: ˆ T 7: U, S, V ← SVD of 8: T ← UV ˆ T )/Tr ( T ) 9: k ← Tr (T T ˆ 2F < ||kopt Topt − ˆ opt ||2F then 10: if ||kT − || ˆ opt ← ˆ 11: 12: end if 13: end while 14: end while gradually converge to a local optimum; thus, the increase in search width will compare numerous local optimal solutions and approximate the global optimum.
8.5 Row generation algorithm The formulated problem in (8.5) is a mixed-integer nonlinear programming (MINLP) problem, as it has both nonlinear orthogonal constraint T T · T = I and the integer ˆ ∈ {−1, 1}m×n . Although such MINLP problem can be solved by existing constraint algorithms such as genetic algorithm [295], it lacks efficiency and only problem in small size can be managed. For the embedding matrix in compressive sensing, ˆ may have the transformation matrix T could have dozens of rows while matrix thousands of Boolean variables, so current solvers may fail in such scale. In this section, we proposed a row generation algorithm that also can efficiently tackle the problem.
8.5.1 Elimination of norm equality constraint The orthonormality of T in (8.5) implies two specific constraints, the orthogonality of rows of T that tiT · tj = 0, ∀ i, j that i = j
(8.16)
and the norm equality that ti 22 = 1, ∀ i
(8.17)
Large-scale case study: accelerator for compressive sensing
201
where ti is the ith row of T . Both imply numerous quadratic equality constraints (nonconvex) and therefore it is hard to manage simultaneously. The nonconvex quadratic norm equality constraint of rows of T indicates the normalization of rows after orthogonality is satisfied. In the following, we show how the norm equality constraint can be eliminated without affecting the solution accuracy of problem in (8.5). Assume we only impose orthogonal constraint on T rather than more strict orthonormal constraint, the original problem can be then relaxed to ˆ 2F minimize T − ˆ T ,
subject to T T · T = D2 ˆ ∈ {−1, 1}m×n
(8.18)
where D = diag(d1 , d2 , . . . , dm ) is a diagonal matrix, and di is the norm of ith row of T . That is to say, an additional row scaling operation is introduced during the sensing stage ˆ y = D−1 x
(8.19)
ˆ T is the optimized Boolean embedding matrix that can be efficiently where realized in hardware, is the orthonormal sparse basis of original signal and x is the sparse coefficient. In fact, the row scaling operation during signal acquisition is unnecessary and can be transferred to recovery stage if an implicit sensing is performed, ˆ yˆ = x
(8.20)
with corresponding signal reconstruction by minimize x1 x∈RN
subject to |D−1 yˆ − D−1 T x| ≤ ε
(8.21)
where ε is the tolerance for noise on sampled signal data yˆ . As such, the norm equality constraint is eliminated while the compressive sensing signal acquisition front-end hardware complexity stays the same and recovery quality is not affected.
8.5.2 Convex relaxation of orthogonal constraint To construct a transformation matrix T with orthogonal rows and minimize the cost function at the same time is challenging, in the following, we propose a convex row generation algorithm that seeks local optimal solution. The idea is to construct each
202
ReRAM-based machine learning
row of T at one time while minimizing the cost function. Assume t1 , t2 , . . . , ti−1 are first i − 1 rows that are already built with orthogonality, to construct the ith row ti , minimize ti − ψˆ i 22 ti ,ψˆ i
⎡
t1 ⎢t2 ⎢ subject to ⎢. ⎣..
⎤ ⎥ ⎥ T ⎥ · ti = 0 ⎦
ti−1 ψˆ i ∈ {−1, 1}n
(8.22)
In other words, each time to construct a new row ti , it has to be orthogonal with previously built t1 , t2 , . . . , ti−1 . The iterative row generation algorithm is shown in Algorithm 8.2. From the geometric perspective, Algorithm 8.2 seeks to find an orthogonal basis in the m-dimensional space iteratively. Initially T is empty so the first direction vector has the greatest freedom to minimize the cost function. After the first basis vector is chosen, the algorithm finds the best basis vector in the left (m − 1)-dimensional subspace that best minimizes the target function. This is iteratively performed until the last direction is selected in only one-dimensional subspace with freedom for length only. As T is a square matrix, there always exists a solution for Algorithm 8.2. The MINLP problem with m × n integer variables in (8.18) is therefore relaxed to m MINLP subproblems integer programming problems each with only n integer variables.
8.5.3 Overall optimization algorithm The overall algorithm to solve (8.18) is illustrated in Algorithm 8.2. The 0-1 programming problem (8.22) within the loop can be readily solved by branch-and-cut method, under the condition that the number of Boolean variable is kept small. The brand-and-cut method is widely implemented in solvers such as MOSEK [296] and BARON [297].
Algorithm 8.2: Iterative row generation algorithm Input: real-valued embedding matrix Output: orthogonal transformation matrix T , optimized Boolean embedding ˆ matrix ˆ =∅ 1: Initialize T = ∅, 2: for i ← 1 to m do 3: get ti by solving in (8.22) problem ˆ T ˆ 4: , = ˆ update T = ti ψi 5: end for
Large-scale case study: accelerator for compressive sensing
203
Without the linearization by row generation, the branch-and-cut method cannot be applied as the orthogonal constraint is strongly nonlinear and thus the evaluation of lower and upper bounds for each subproblem will be extremely complicated. In addition, the linearization by row generation significantly reduces the number of Boolean variables and thus reduces the worst-case complexity from 2m×n to m · 2n . As such, the row generation together with widely available integer programming solvers can find solution for problem formulated in (8.18).
8.6 Numerical results 8.6.1 Experiment setup In this part, we evaluate different compressive sensing sampling matrices from both software and hardware perspectives. The numerical experiments are performed within MATLAB® on a desktop with 3.6 GHz Intel i7 processor and 16 GB memory. The software performance of sampling matrices is mainly characterized by the signal recovery quality of sampling matrices. For this purpose, both labeled faces in the wild (LFW) image data [298] and biomedical electrocardiogram (ECG) data [299] are used. For both types of data, the NuMax [289] optimization is first applied with varied training parameter δ values ([0.05, 0.1, . . . , 0.35]), NuMax produces optimized real-valued sampling matrices with different ranks. As depicted in the flowchart in Figure 8.2, the proposed algorithms are then applied to Booleanize NuMax sampling matrices. Apart from the above data-driven sampling matrices, random Gaussian, Bernoulli and Reed–Muller [287,288] (non-data-driven optimization) sampling are also compared. The RSNR is used as signal recovery quality metric, which is defined as x2 RSNR = 20 log10 (8.23) x − xˆ 2 where x is the original signal and xˆ is the reconstructed signal. With respect to hardware cost consideration, above all sampling matrices can be mapped to three different sampling hardware configurations. Specifically, the MAC/SRAM, MAC/PRNG and ReRAM crossbar configurations with their variations are evaluated to examine the hardware friendliness of all the sampling schemes. For real-valued MAC, 16-bit resolution is used as we find that resolution higher than 16-bit will not improve accuracy, as shown in Figure 8.1. For ReRAM crossbar, the resistance of 1K and 1M are used for ReRAM on-state resistance and offstate resistance according to [293]. The area of the ReRAM crossbar is evaluated by multiplying the cell area (4F 2 ) with sampling matrix size plus one additional column to calculate current offset as discussed in Section 8.3.2. Dynamic power of the ReRAM crossbar is evaluated statistically under 1,000 random input patterns following an uniform distribution with voltage ranging from −0.5 to 0.5 V (|V | < |Vset | = 0.8 V and |V | < |Vreset | = 0.6 V [293]) and the duration of operation is 5 ns [293]. Both the real-valued and Boolean digital CMOS matrix multiplier designs are implemented in Verilog and synthesized with GlobalFoundries 65-nm low-power process design kit
204
ReRAM-based machine learning
(PDK). A PRNG design [286] is also implemented and synthesized. The SRAM that stores the sampling matrix is evaluated by CACTI [300] memory modeling tool with 65-nm low standby power process opted. Table 8.3 shows all valid combinations of sampling matrix and hardware configuration that will be compared in the section. Among all the combinations, we will show in this section that our proposed sampling matrix can achieve both the best signal recovery quality and hardware efficiency.
8.6.2 IH algorithm on high-D ECG signals For training stage (NuMax), 1,000 ECG periods in dimension of 256 are randomly picked from database [299] as the data set χ , which leads to around 1 million of pairwise distance vectors in set S(χ ). For testing phase, another 1,000 ECG periods are selected as data set χ , which have no overlap with learning data set χ . The ECG signal reconstruction is performed on unseen data set χ by solving (8.1) with Battle–Lemarie wavelet bases used.
8.6.2.1 Algorithm convergence and effectiveness The efficiency of Algorithm 8.1 can be examined from two aspects, finding both local and global optima. The efficiency of finding local optimum is assessed by convergence ˆ 2F stops rate. The local search terminates when the approximation error ||T − || improving. Given specific RIP upper bounds, NuMax [289] provides with different ranks. With RIP constraint of 0.1, the NuMax produces a ∈ R19×256 sampling matrix. Algorithm 8.1 is applied to with total 10,000 repeated local search and the convergence is illustrated in Figure 8.6(a). It can be observed that the relative error reduces dramatically within the first few iterations. The zoomed subfigure shows that local search on average converges within 50 iterations, where convergence is defined as less than 1e-6 error reduction in two consecutive iterations. Generally, the local optimum can be considered found in less than 100 iterations. The global search is achieved by scattering many initial guesses in the orthogonal matrices space for T , and comparing the corresponding local optima. The errors under varying number of initial guesses are shown in Figure 8.6(b). Considering the Boolean Table 8.3 Summary of the embedding methods to be compared Embeddings
Boolean?
Optimized?
Construction
NuMax This technique Genetic algorithm Random Boolean Gaussian
× ×
× ×
opt. on χ opt. on NuMax† opt. on NuMax‡ randomize randomize
† ‡
Solve (8.5). Solve (8.5) by genetic algorithm from [295].
Large-scale case study: accelerator for compressive sensing 90
10
70 60 50
0.08 0.06 0.04 0.02
8 Convergence iterations
4
0 0
40
6
Distribution
50
100
2
150
RSNR (dB)
Probability
80 MSE
205
0
30 1
10
20
30 Iteration
40
50
-2
24.6
10.2
24.1
Information loss Recovery quality 9.95
23.6 100
101 102 Number of guesses
RSNR (dB)
MSE
(a)
9.7 103
(b)
Figure 8.6 The algorithm efficiency for (a) local search convergence and (b) global search convergence
constraint and the orthogonal constraint, the problem formulated in (8.5) is generally NP-hard. Therefore, the relative error can be improved by scattering exponentially more initial guesses, yet no convergence is observed. Hence, an efficient search policy should be designed in a way scattering as many initial points as possible and limiting the local search for each initial guess within 100 iterations.
8.6.2.2 ECG recovery quality comparison 19 The ECG signal recovery examples under γ = 256 are shown in Figure 8.7. For nondata-driven sampling matrices, both the random Bernoulli and Gaussian show similar reconstruction quality. In other words, the increase of bits in random numbers will not improve recovery quality. This is because, the increase of bits of random number will not gain any additional information. The PRNG-based Bernoulli exhibits the lowest reconstructed signal quality. This is because, as PRNG produces 0/1 sequences with a predetermined pattern, it has self-coherence issue. For example, an 8-bit per cycle PRNG has a period of 256 (28 ), and when filling a sampling matrix by rows with such 0/1 sequences, all rows of the matrix will be identical. The Reed–Muller code optimizes the sampling matrix by minimizing the correlations among different rows/columns, which helps to improve sampling performance.
ReRAM-based machine learning
Amplitude
206
Original ECG signal Reconstructed (PRNG)
200 100 0 –100 200 100 0 –100 200 100 0 –100 200 100 0 –100 200 100 0 –100 200 100 0 –100 200 100 0 –100
Gaussian
Bernoulli
Reed–Muller
Proposed (Lv2)
Proposed (Lv3)
NuMax
100
200
300
400
500
600
700
800
900
1,000
Time
(a) 20
RSNR (dB)
15 10
3.3x
5
NuMax Proposed (Lv2) Proposed (Lv3) Reed–Muller Gaussian Bernoulli PRNG
9db
0 –5 0 (b)
0.05
0.1
0.15 0.2 0.25 Undersampling ratio γ
0.3
0.35
0.4
Figure 8.7 The recovery quality comparison among different sampling matrices: (a) examples of recovered ECG signals at γ = 19/256 and (b) RSNR for 1,000 ECG periods
Being a generic sampling matrix that works with all data types, it cannot exploit the isometric property of ECG signal, which limits its performance in particular applications. Specifically, it only shows 1 dB improvement compared to random Bernoulli sampling.
Large-scale case study: accelerator for compressive sensing
207
The data-driven NuMax real-valued sampling exhibits the best recovery quality (highest RSNR) as shown in Figure 8.7(b). The proposed iterative heuristic (IH) algorithm quantizes the real-valued NuMax sampling with slight quality loss. Specifically, 19 at the undersampling ratio γ = 256 , the IH (lv2) exhibits 8, 9 and 10 dB higher RSNR than that of Reed–Muller, Bernoulli and pseudo-Bernoulli samplings, respectively. Also, Level 3 quantization through (8.15) can preserve more information than Level 2 quantization through (8.13). The RSNR of IH (lv3) shows marginal 0.48 dB higher RSNR than IH (lv2). The IH (lv3) sampling matrix ∈ {−1, 0, 1}m×n will incur additional hardware overhead compared to IH (lv2) ∈ {−1, 1}m×n . Figure 8.7(a) gives a visual effect of quality of recovered ECG signal segments with different sampling matrices. The data-driven sampling matrices, i.e., NuMax, IH (lv2) and IH (lv3), can recover signals that tightly coincide with original signals.
8.6.3 Row generation algorithm on low-D image patches For training stage (NuMax), 6,000 patches with size of 8 × 8 are randomly picked throughout all images as the data set χ , which leads to around 18 million of pairwise distance vectors in set S(χ ). For testing phase, another 6,000 patches with size of 8 × 8 are selected as data set χ with no overlap with learning data set χ . The image reconstruction is performed on unseen data set χ by solving (8.1) with 2D discrete cosine transform (DCT) bases. The genetic algorithm [295] is adopted as the baseline solver for the MINLP in (8.5), which is compared with the proposed algorithm in Algorithm 8.3. Both algorithms are run given same amount of time, i.e., m × 500 s where m is the rank of that indicates the size of the problem.
8.6.3.1 Algorithm effectiveness The idea behind the proposed real-valued matrix Booleanization is to preserve the RIP of NuMax sampling matrix, which differs from the truncation-based quantization. The information loss during the quantization is directly related to the RIP preservation. The algorithm effectiveness in this part will be examined by the isometric distortion δ, defined in (8.2). The distortions of all embeddings are tested on unseen data set χ (see Figure 8.8). The isometric distortions of both random embeddings are almost invariant. Being optimized on image data set χ , both the NuMax and proposed (quantized NuMax) are significantly better than random embeddings. With focus on the Boolean sampling matrices that are hardware friendly, the isometric distortion of optimized Boolean embedding is 3.0× better than random Boolean embedding on average. Due to the near-orthogonal rotation, the optimized Boolean embedding experiences some penalty on isometric distortion δ compared to NuMax approach. For genetic algorithm, as it experiences higher distortion than that of the proposed algorithm, it can be inferred that Algorithm 8.3 can find a more precise solution. In addition, it can be observed that the genetic algorithm fails when undersampling ratio mn increases, and this is because the proposed row generation-based algorithm
208
ReRAM-based machine learning 0.7 NuMax Proposed
Isometric distortion δ
0.6
Genetic algorithm
0.5
Gaussian Random Boolean
0.4 0.3 0.2 0.1 0 0.4
0.5
0.6 0.7 0.8 Undersampling ratio γ
0.9
1
Figure 8.8 The isometric distortion on the unseen data set χ for different embeddings Algorithm 8.3: Iterative row generation algorithm Input: real-valued embedding matrix Output: orthogonal transformation matrix T , optimized Boolean embedding ˆ matrix ˆ =∅ 1: initialize T = ∅, 2: for i ← 1 to m do 3: while U − L > ε do 4: for all branch with feasible set S do 5: partition the feasible set S into S1 and S2 6: evaluate L (S1 ), L (S2 ) 7: update lower bound L = min{L , L (S1 ), L (S2 )} 8: evaluate U (S1 ), U (S2 ) 9: update upper bound U = min{U , U (S1 ), U (S2 )} 10: if L (S1 ) or L (S2 ) > U then 11: cut branch S1 or S2 12: end if 13: end for 14: end while ˆ T ˆ , = ˆ 15: update T = ti ψi 16: end for
requires linearly more time when the number of row m increases, while the genetic algorithm needs exponentially more time. Moreover, the solution provided by genetic algorithm is stochastic, which has no guarantee on its effectiveness while the proposed algorithm is deterministic.
Large-scale case study: accelerator for compressive sensing
209
8.6.3.2 Image recovery quality comparison The recovery examples under γ = 25 are shown in Figure 8.9(a). The reconstructed 64 images in blue box correspond to Boolean embeddings that have low-power hardware implementations, and images in red box are from optimization-based approaches
Original image
NuMax
Reed– Muller
Proposed
Random Boolean
Genetic algorithm
PRNG
Gaussian
Low power
Low error (a) 40 35
RSNR (dB)
30 25 20 NuMax Proposed Genetic algorithm Reed–Muller Gaussian Bernoulli PRNG
1.6x
15 8.3dB 10 5 0 0.3 (b)
0.4
0.5
0.6
0.7
0.8
0.9
Undersampling ration γ
Figure 8.9 The recovery quality comparison among different embedding matrices: (a) examples of recovered images under γ = 25/64 and (b) RSNR on 6,000 8 × 8 image patches
210
ReRAM-based machine learning
which show lower recovery errors. The genetic algorithm is also optimization based, but the effectiveness is inconsiderable. Therefore, only the proposed can achieve both low power and high recovery performance. The numerical image reconstruction quality is shown in Figure 8.9(b). The two random embeddings show similar reconstruction RSNR, which is averagely 8.3 dB lower than that of the proposed optimized Boolean sampling matrix. The RSNR of optimized Boolean embedding is close to that of NuMax embedding, which is 2.5 dB lower as a result of information loss by near-orthogonal rotation. On the other hand, the genetic algorithm shows no obvious effectiveness of improving recovery quality even though it optimizes a Boolean embedding matrix. ˆ too much information loss The main reason is that during the conversion of to ˆ to be close to a random Boolean matrix. In other words, the genetic algorithm leads is ineffective to solve the problem in (8.5). In addition, the stochastic nature of genetic algorithm makes it necessary to perform the algorithm considerably many times. The proposed algorithm, on the contrary, guarantees to produce a Boolean matrix with high performance with single execution.
8.6.4 Hardware performance evaluation In this part, the hardware performance benefits of Boolean embedding will be investigated in detail. The evaluation only focuses on the embedding hardware as indicated by red dash-lined boxes in Figures 8.3 and 8.5.
8.6.4.1 Hardware comparison The matrix–vector multiplier is composed of multiple MACs in parallel. To multiply the signal vector with a 19×256 sampling matrix, 19 MACs are needed and each MAC requires 256 cycles to perform the inner-product with each cycle (1ns). To store the NuMax real-valued sampling matrix, 16kB SRAM with 64-bit I/O bus-width is used. The proposed Boolean optimization quantizes NuMax sampling matrix into a {−1, 1}m×n Boolean matrix. The size of SRAM to store sampling matrix is therefore reduced from 16kB to 1kB. Compared to a Bernoulli {0, 1}m×n matrix, a {−1, 1}m×n multiplication requires calculations of 2’s complement of input signal vector, which incurs additional hardware cost for MACs. To minimize the overhead of 2’s complement, the MAC design in Figure 8.4(c) is used, which calculates 2’s complement only once every 256 cycles. ReRAM crossbar supports both {0, 1}m×n and {−1, 1}m×n Boolean matrices. As the sampling matrix is embedded into the ReRAM crossbar which also performs the matrix multiplication, no separate memory is required. The performance of four hardware schemes that support different types of sampling matrices is compared in Table 8.4. Compared to the NuMax real-valued embedding on 16-bit MAC and 16 kB SRAM hardware, the proposed quantized −1/1 Boolean embedding on 1-bit MAC and 1 kB SRAM consumes 4.6× less operation energy per embedding, 1.8× smaller leakage power and 1.9× smaller area. This is
Large-scale case study: accelerator for compressive sensing
211
Table 8.4 Hardware performance comparison among different sampling matrices (19×256) on varied hardware configurations. Matrix type
Hardware configuration
Energy (nJ)
Leakage power (μW)
Area (μm2 )
Real-valued
MAC (16-bit) MEM (16 kB) MAC (1-bit) MEM (1 kB) MAC (1-bit) PRNG ReRAM crossbar
116.38 8.08 24.81 2.43 21.87 8.26e-2 1.06
119.63 4.66 69.22 0.29 30.40 0.04 —
127,984 31,550 73,207 9,800 29,165 32 173
−1/1 Boolean 0/1 Bernoulli Boolean †
Cycle 256 256 ∼512† 1
PRNG used produces 10 bits per cycle.
because, as mentioned in Section 8.3, the real-valued multiplier generally requires quadratically increasing number of full adders when resolution increases, while Boolean multiplier only needs linearly more full adders. When the proposed quantized −1/1 Boolean embedding is performed on ReRAM crossbar, it further improves the hardware performance significantly. Specifically, for the operation energy per embedding, the ReRAM crossbar-based embedding outperforms the CMOS circuit-based real-valued embedding by 117×. The area of the ReRAM crossbar-based embedding is nearly 1, 000× better than that of CMOS circuit-based real-valued embedding. In addition, the ReRAM crossbar will not experience the leakage power which is at the scale of hundreds of microwatts for the CMOS circuit-based approach. For the operation speed, the ReRAM crossbar embedding executes in single cycle while the CMOS circuit requires 256 cycles due to the reuse of hardware. The overall performance for different sampling matrices on varied hardware platforms is summarized in Table 8.5.
8.6.4.2 Impact of ReRAM variation One nonnegligible issue of mapping Boolean embedding matrix to ReRAM crossbar is the ReRAM HRS and LRS variations. With high resistance variation, the embedding matrix will deviate from expected values to be represented by ReRAM resistance, and hence the recovery quality may degrade. The sensitivity study of recovery quality on the resistance variations of ReRAM is shown in Figure 8.10. The resistance of ReRAM is assumed to follow log-normal distribution with the mean to be RLRS and RHRS , and standard deviation σLRS and σHRS for LRS and HRS cells, respectively. With varied σLRS and σHRS , it can be observed from Figure 8.10 that the performance degradation is more susceptible to resistance variation of LRS, while less sensitive on variation of HRS. In practice, the HRS variation σHRS is approximately 0.3 [301], and LRS variation σLRS roughly 0.1 [301]. The real-world σ is annotated in Figure 8.10 and it can be concluded that the proposed Boolean embedding on
Table 8.5 Comparison of all valid sampling matrices and hardware combinations Platform
Construction
Sampling matrix
Sampling hardware
Boolean?
Optimized?
Data driven?
Gaussian Bernoulli Bernoulli Pseudo-Bernoulli NuMax [289] Reed–Muller [287,288] Reed–Muller [287,288] Proposed† Proposed
MAC(16-bit) + MEM(16kB) MAC(1-bit) + MEM(1kB) ReRAM crossbar MAC(1-bit) + PRNG MAC(16-bit) + MEM(16kB) MAC(1-bit) + MEM(1kB) ReRAM crossbar MAC(1-bit) + MEM(1kB) ReRAM crossbar
× ×
× × × ×
× × × × × ×
‡
Stage
Effort
Recovery quality‡
Energy consumed
Off-line Off-line Off-line Runtime Off-line Off-line Off-line Off-line Off-line
Immediate Immediate Immediate — ∼100s Fast Fast ∼100s ∼100s
1.12 1.09 1.09 1.00 3.86 1.25 1.25 3.18 3.18
117.4 22.9 1.0 20.7 117.4 22.9 22.9 25.7 1.0
The recovery quality depicts the quality performance ratio of all sampling matrices over baseline pseudo-Bernoulli. The RSNR dB for ECG is converted to mean-squared error. Numbers in bold are ones with good performance. The energy consumption is shown as ratio of used energy of all sampling matrices over that of ReRAM crossbar. Numbers in bold are ones with good performance. † The proposed denotes the Booleanized NuMax sampling matrix by proposed Algorithms.
Large-scale case study: accelerator for compressive sensing 40
40
Real-world σLRS≈0.1
30 RSNR (dB)
213
30 20
25×64 27×64 28×64 30×64 34×64 39×64 48×64
20 10 Real-world σHRS≈0.3 10
0 0
1
(a) 15
2 3 σHRS
4
0
5
0.1
0.2 0.3 σLRS
0.4
0.5
0.4
0.5
14
Real-world σHRS≈0.3
12
RSNR (dB)
10 10 5 31×256 29×256 24×256 22×256 19×256 17×256 15×256
0
8
6 Real-world σLRS≈0.1
-5
4 0
(b)
1
2 3 σHRS
4
5
0
0.1
0.2 0.3 σLRS
Figure 8.10 The sensitivity of recovery quality of (a) image signal and (b) ECG signal on the resistance standard deviation σ of ReRAM for both LRS and HRS following log-normal distribution
ReRAM crossbar is robust against ReRAM device variations when on/off ratio is high (GLRS GHRS ≈ 0). To further suppress the performance degradation, material engineering [302] and verification programming method [301] can help achieve higher LRS uniformity.
This page intentionally left blank
Chapter 9
Conclusions: wrap-up, open questions and challenges
9.1 Conclusion To conclude, this book has shown a thorough study on resistive random-access memory (ReRAM)-based nonvolatile in-memory architecture towards machine learning applications from circuit level, to architecture level, and all the way to system level. For the circuit level, on the one hand, we use non-volatile memory (NVM)-SPICE to simulate matrix–vector multiplication acceleration by binary ReRAM crossbar. By programming all the elements of a matrix into ReRAM resistance in the crossbar array, we can use crossbar wordline voltages to represent another matrix, and the multiplication result can be denoted by the merging current on bitlines. In this mapping scheme, both ReRAM resistance and crossbar I/O are in binary format so that the computation accuracy can be guaranteed. Simulation results show that the proposed architecture has shown 2.86× faster speed, 154× better energy efficiency and 100× smaller area when compared to the same design by complimentary metal oxide semiconductor (CMOS)-based application specific integrated circuit (ASIC). On the other hand, we use a Verilog-A model to build an ReRAM-based voltage-controlled oscillator circuit in Cadence. In a coupled network by ReRAM-based oscillator, the simulation results can be fitted as L2-norm calculation. Compared to traditional CMOS-based oscillator circuit, it has a much simpler structure with smaller area and better energy efficiency. For the architecture level, we have developed a distributed in-memory architecture (XIMA) and three-dimensional (3D) CMOS-ReRAM architecture. For XIMA, we use both ReRAM crossbar for data storage and computing engine. We use a data block and logic block to form a data–logic pair, and all the pairs are located distributively. As a result, the communication bandwidth can be significantly improved because the computing engine only accesses the data from pairs. To achieve the in-memory computing, we design a control bus in each data–logic pair for communication among processor, data and logic block. In addition, communication protocol between processor and control bus is redefined for the in-memory computing implementation. For 3D CMOS-ReRAM architecture, we developed two schemes: single-layer and multilayer. The single-layer architecture uses ReRAM via connecting the top-layer wordlines and bottom-layer bitlines. All the other CMOS logics are implemented in the bottom layer as well. The multilayer architecture uses through
216
ReRAM-based machine learning
silicon via (TSV) connecting the first-layer ReRAM-based data buffer, the secondlayer ReRAM-based logic computing and the third-layer CMOS. Such a 3D architecture can significantly improve the throughput as well as the area efficiency for the hybrid ReRAM/CMOS computing system. For the system level, for the XIMA, we have accelerated three machine learning algorithms. The learning and inference procedures of single-layer feedforward neural network (SLFN) have been optimized and partially mapped on the passive binary ReRAM crossbar. In addition, we mapped the binary convolutional neural network on both passive array and One Selector One ReRAM array with different mapping schemes. Moreover, L2-norm gradient-based learning and inference are also implemented on an ReRAM network with both crossbar and coupled oscillators. For 3D CMOS-ReRAM architecture, we also mapped the optimized SLFN algorithm. All the operations in learning and inference stages are implemented so that it can achieve the online learning. In addition, tensorized neural network is mapped on both single-layer and multilayer accelerators with different mapping scheme. For all the machine learning algorithms mapped on XIMA and 3D CMOS-ReRAM accelerator, we evaluate their performance in device, architecture and system levels. All the implementations show higher throughput, bandwidth and parallelism with better energy efficiency.
9.2 Future work Based on the above works, there are a few potential future works to enhance this research. The first potential work is to validate some designs discussed. In this work, we explore the binary ReRAM crossbar as well as oscillator network in device and circuit level by simulation. However, the ReRAM devices in the array are not fully uniform so that the mapping scheme may not be applied in a large crossbar array. In addition, sneak path and thermal effect in the ReRAM crossbar will also occur the inaccuracy during computing. As for the XIMA, this work only simulates a few groups of control bus design, but in the real implementation, there will be thousands of control bus blocks so it is also necessary to optimize all these designs. The second potential work is to realize neuromorphic computing by the ReRAM network. This work only explores the artificial neural network implementations in the ReRAM-based architecture. All of these algorithms are originally designed for CMOS-based computing systems such as CPU, GPU, CMOS-ASIC and FPGA. The ReRAM-based architecture can perform analog-value computing operations, which is widely applied in neuromorphic applications. For example, it is easier to implement spike-timing-dependent plasticity-based synapse programming algorithm and also spiking neural networks on ReRAM compared to CMOS-ASIC or FPGA. It is possible to explore an ReRAM-based neuromorphic computing system with simpler structure and better scalability than the state-of-the-art works.
References
[1] [2] [3] [4]
[5] [6] [7]
[8]
[9] [10]
[11] [12]
[13] [14]
[15]
Cukier K. Data, data everywhere. The Economist. 2010 Feb. International Data Corporation. Digital Universe Study. Dell EMC. 2011. White T. O’Reilly Media Inc; 2012. Zaharia M, Chowdhury M, Franklin MJ, et al. Spark: Cluster computing with working sets. In: USENIX Conference on Hot Topics in Cloud Computing; 2010. Apache Flink. Apache Flink: Scalable batch and stream data processing. http://flink.apache.org/. Online. Carbone P, Fóra G, Ewen S, et al. Lightweight asynchronous snapshots for distributed dataflows. CoRR. 2015;abs/1506.08603. Saha B, Shah H, Seth S, et al. Apache Tez: A unifying framework for modeling and building data processing applications. In: ACM SIGMOD International Conference on Management of Data; 2015. Akidau T, Bradshaw R, Chambers C, et al. The Dataflow model: A practical approach to balancing correctness, latency, and cost in massive-scale, unbounded, out-of-order data processing. Proc VLDB Endow. 2015;8(12): 1792–1803. LeDell E. High performance machine learning in R with H2O. In: ISM HPC on R Workshop; 2015. Xing EP, Ho Q, Dai W, et al. Petuum: A new platform for distributed machine learning on big data. In: ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; 2015. Bakshi K. Considerations for big data: Architecture and approach. In: IEEE Aerospace Conference; 2012. Chen Y, Alspaugh S and Katz R. Interactive analytical processing in big data systems: A cross-industry study of MapReduce workloads. Proc VLDB Endow. 2012;5(12):1802–1813. Lee KH, Lee YJ, Choi H, et al. Parallel data processing with MapReduce: A survey. SIGMOD Rec. 2012;40(4):11–20. Shvachko K, Kuang H, Radia S, et al. The Hadoop distributed file system. In: IEEE Symposium on Mass Storage Systems and Technologies (MSST); 2010. Magaki I, Khazraee M, Gutierrez LV, et al. ASIC clouds: Specializing the data center. In: ACM/IEEE International Symposium on Computer Architecture (ISCA); 2016.
218 [16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
ReRAM-based machine learning Caulfield AM, Chung ES, Putnam A, et al. A cloud-scale acceleration architecture. In: IEEE/ACM International Symposium on Microarchitecture (MICRO); 2016. Gutierrez A, Cieslak M, Giridhar B, et al. Integrated 3D-stacked server designs for increasing physical density of key-value stores. SIGPLAN Not. 2014;49(4):485–498. Putnam A, Caulfield AM, Chung ES, et al. A reconfigurable fabric for accelerating large-scale data center services. SIGARCH Comput Archit News. 2014;42(3):13–24. Wess M, Manoj PDS and Jantsch A. Weighted quantization-regularization in DNNs for weight memory minimization towards HW implementation. IEEE Transactions on Computer Aided Systems of Integrated Circuits and Systems. 2018. Lechner M, Jantsch A and Manoj PDS. ResCoNN: Resource-efficient FPGAaccelerated CNN for traffic sign classification. In: IEEE International Green and Sustainable Computing Conference (IGSC); 2019. Vashist A, Keats A, Manoj PDS, et al. Unified testing and security framework for wireless network-on-chip enabled multi-core chips. In: ACM/IEEE International Conference on Hardware/Software Codesign and System Synthesis (CODES+ISSS); 2019. Wess M, Manoj PDS and Jantsch A. Weighted quantization-regularization in DNNs for weight memory minimization towards HW implementation. In: International Conference on Hardware/Software Codesign and System Synthesis (CODES+ISSS); 2018. Yu H, Ni L and Wang Y. Non-volatile In-Memory Computing by Spintronics. Vol. 2. Morgan & Claypool Publishers; 2016. Available from: https://dl.acm.org/doi/book/10.5555/3086882. Yu H and Wang Y. Design Exploration of Emerging Nano-scale Nonvolatile Memory. Springer; 2014. Available from: https://www.springer. com/gp/book/9781493905508. Song Y, Yu H and Manoj PDS. Reachability-based robustness verification and optimization of SRAM dynamic stability under process variations. IEEE Trans on Computer-Aided Design of Integrated Circuits and Systems. 2014;33(4):585–598. Song Y, Manoj PDS and Yu H. Zonotope-based nonlinear model order reduction for fast performance bound analysis of analog circuits with multipleinterval-valued parameter variations. In: Design, Automation Test in Europe Conference Exhibition (DATE); 2014. Song Y, Manoj PDS and Yu H. A robustness optimization of SRAM dynamic stability by sensitivity-based reachability analysis. In: Asia and South Pacific Design Automation Conference (ASP-DAC); 2014. Wu SS, Wang K, Manoj PDS, et al. A thermal resilient integration of manycore microprocessors and main memory by 2.5D TSI I/Os. In: Design, Automation Test in Europe Conference Exhibition (DATE); 2014.
References [29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
219
Burger D, Goodman JR and Kägi A. Memory Bandwidth Limitations of Future Microprocessors. SIGARCH Comput Archit News. 1996;24(2): 78–89. Manoj PDS, Yu H, Gu C, et al. A zonotoped macromodeling for reachability verification of eye-diagram in high-speed I/O links with jitter. In: IEEE/ACM International Conference on Computer-Aided Design (ICCAD); 2014. Makrani HM, Sayadi H, Manoj PDS, et al. A comprehensive memory analysis of data intensive workloads on server class architecture. In: International Symposium on Memory Subsystems; 2018. Makrani HM, Sayadi H, Manoj PDS, et al. Compressive sensing on storage data: An effective solution to alleviate I/O bottleneck in data-intensive workloads. In: IEEE International Conference on Application-specific Systems, Architectures and Processors; 2018. Manoj PDS, Wang K, Huang H, et al. Smart I/Os: A data-pattern aware 2.5D interconnect with space-time multiplexing. In: ACM/IEEE International Workshop on System Level Interconnect Prediction (SLIP); 2015. Hantao H, Manoj PDS, Xu D, et al. Reinforcement learning based selfadaptive voltage-swing adjustment of 2.5D I/Os for many-core microprocessor and memory communication. In: IEEE/ACM Int. Conf. on Computer-Aided Design (ICCAD); 2014. Xu D, Manoj PDS, Huang H, et al. An energy-efficient 2.5D through-silicon interposer I/O with self-adaptive adjustment of output-voltage swing. In: IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED); 2014. Wang J, Ma S, Manoj PDS, et al. High-speed and low-power 2.5D I/O circuits for memory-logic-integration by through-silicon interposer. In: IEEE International 3D Systems Integration Conference (3DIC); 2013. Manoj PDS, Wang K andYu H. Peak power reduction and workload balancing by space-time multiplexing based demand-supply matching for 3D thousandcore microprocessor. In: ACM/EDAC/IEEE Design Automation Conf.; 2013. Manoj PDS and Yu H. Cyber-physical management for heterogeneously integrated 3D thousand-core on-chip microprocessor. In: IEEE International Symposium on Circuits and Systems (ISCAS); 2013. Manoj PDS, Joseph A, Haridass A, et al. Application and thermal reliabilityaware reinforcement learning based multi-core power management. ACM Transactions on Embedded Computing Systems. 2019. Pagani S, Manoj PDS, Jantsch A, et al. Machine learning for power, energy, and thermal management on multi-core processors: A survey. IEEE Transactions on Computer Aided Systems of Integrated Circuits and Systems. 2018. Manoj PDS, Jantsch A and Shafique M. SmartDPM: Dynamic power management using machine learning for multi-core microprocessors. Journal of Low-Power Electronics. 2018.
220 [42]
[43]
[44]
[45] [46]
[47]
[48] [49] [50] [51]
[52]
[53] [54]
[55]
[56]
ReRAM-based machine learning Manoj PDS, Lin J, Zhu S, et al. A scalable network-on-chip microprocessor with 2.5D integrated memory and accelerator. IEEE Transactions on Circuits and Systems I: Regular Papers. 2017;64(6):1432–1443. Matsunaga S, Hayakawa J, Ikeda S, et al. MTJ-based nonvolatile logicin-memory circuit, future prospects and issues. In: Proceedings of the Conference on Design, Automation and Test in Europe. European Design and Automation Association; 2009. pp. 433–435. Matsunaga S, Hayakawa J, Ikeda S, et al. Fabrication of a nonvolatile full adder based on logic-in-memory architecture using magnetic tunnel junctions. Applied Physics Express. 2008;1(9):091301. Kautz WH. Cellular logic-in-memory arrays. IEEE Transactions on Computers. 1969;100(8):719–727. Kimura H, Hanyu T, Kameyama M, et al. Complementary ferroelectriccapacitor logic for low-power logic-in-memory VLSI. IEEE Journal of Solid-State Circuits. 2004;39(6):919–926. Hanyu T, Teranishi K and Kameyama M. Multiple-valued logic-in-memory VLSI based on a floating-gate-MOS pass-transistor network. In: SolidState Circuits Conference, 1998. Digest of Technical Papers. 1998 IEEE International. IEEE; 1998. pp. 194–195. Kouzes RT, Anderson GA, Elbert ST, et al. The changing paradigm of dataintensive computing. Computer. 2009;(1):26–34. Wolpert DH. The lack of a priori distinctions between learning algorithms. Neural Computation. 1996;8(7):1341–1390. Hinton GE, Osindero S and Teh YW. A fast learning algorithm for deep belief nets. Neural Computation. 2006;18(7):1527–1554. Müller KR, Tangermann M, Dornhege G, et al. Machine learning for realtime single-trial EEG-analysis: From brain–computer interfacing to mental state monitoring. Journal of Neuroscience Methods. 2008;167(1):82–90. Glorot X and BengioY. Understanding the difficulty of training deep feedforward neural networks. In: International Conference on Artificial Intelligence and Statistics; 2010. pp. 249–256. Huang GB, Zhu QY and Siew CK. Extreme learning machine: Theory and applications. Neurocomputing. 2006;70(1):489–501. Coates A, Ng AY and Lee H. An analysis of single-layer networks in unsupervised feature learning. In: International Conference on Artificial Intelligence and Statistics; 2011. pp. 215–223. Park S, Qazi M, Peh LS, et al. 40.4 fJ/bit/mm low-swing on-chip signaling with self-resetting logic repeaters embedded within a mesh NoC in 45nm SOI CMOS. In: Proceedings of the Conference on Design, Automation and Test in Europe. EDA Consortium; 2013. pp. 1637–1642. Kumar V, Sharma R, Uzunlar E, et al. Airgap interconnects: Modeling, optimization, and benchmarking for backplane, PCB, and interposer applications. IEEE Transactions on Components, Packaging and Manufacturing Technology. 2014;4(8):1335–1346.
References [57]
[58] [59]
[60] [61] [62] [63]
[64]
[65]
[66]
[67]
[68]
[69] [70] [71]
[72] [73]
221
Wang Y, Yu H, Ni L, et al. An energy-efficient nonvolatile in-memory computing architecture for extreme learning machine by domain-wall nanowire devices. IEEE Transactions on Nanotechnology. 2015;14(6):998–1012. Akinaga H and Shima H. Resistive random access memory (ReRAM) based on metal oxides. Proceedings of the IEEE. 2010;98(12):2237–2251. Kim KH, Gaba S, Wheeler D, et al. A functional hybrid memristor crossbararray/CMOS system for data storage and neuromorphic applications. Nano Letters. 2011;12(1):389–395. Chua LO. Memristor—The missing circuit element. IEEE Transactions on Circuit Theory. 1971;18(5):507–519. Williams SR. How we found the missing memristor. IEEE Spectrum. 2008;45(12):28–35. Strukov DB, Snider GS, Stewart DR, et al. The missing memristor found. Nature. 2008;453(7191):80–83. Shang Y, Fei W and Yu H. Analysis and modeling of internal state variables for dynamic effects of nonvolatile memory devices. IEEE Transactions on Circuits and Systems I: Regular Papers. 2012;59(9):1906–1918. Fei W,Yu H, Zhang W, et al. Design exploration of hybrid CMOS and memristor circuit by new modified nodal analysis. IEEE Transactions on Very Large Scale Integration (VLSI) Systems. 2012;20(6):1012–1025. Liu X, Mao M, Liu B, et al. RENO: A high-efficient reconfigurable neuromorphic computing accelerator design. In: Proceedings of the 52nd Annual Design Automation Conference. ACM; 2015. pp. 66.1–66.6. Kim Y, Zhang Y and Li P. A digital neuromorphic VLSI architecture with memristor crossbar synaptic array for machine learning. In: International SOC Conference (SOCC). IEEE; 2012. pp. 328–333. Lu W, Kim KH, Chang T, et al. Two-terminal resistive switches (memristors) for memory and logic applications. In: Asia and South Pacific Design Automation Conference (ASP-DAC). IEEE; 2011. pp. 217–223. Dennard RH, Gaensslen FH, Yu H, et al. Design of ion-implanted MOSFET’s with very small physical dimensions. IEEE Journal of Solid-State Circuits. 1974;9(5):256–268. Loh GH. 3D-Stacked memory architectures for multi-core processors. In: International Symposium on Computer Architecture; 2008. Strenz R. Embedded flash technologies and their applications: Status outlook. In: International Electron Devices Meeting; 2011. Energy Limits to the Computational Power of the Human Brain. Last Accessed: June, 15, 2020. Available from: https://foresight.org/Updates/ Update06/Update06.1.php. Laughlin SB, de Ruyter van Steveninck RR and Anderson JC. The metabolic cost of neural information. Nature Neuroscience. 1998;1:36–41. Jiang Z, Wu Y, Yu S, et al. A compact model for metal-oxide resistive random access memory with experiment verification. IEEE Transactions on Electron Devices. 2016;63(5):1884–1892.
222 [74]
[75]
[76]
[77] [78] [79] [80] [81] [82] [83]
[84]
[85]
[86]
[87]
[88]
[89]
[90]
ReRAM-based machine learning Merolla PA, Arthur JV, Alvarez-Icaza R, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science. 2014;345(6197):668–673. Davies M, Srinivasa N, Lin T, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro. 2018;38(1): 82–99. Hu M, Graves CE, Li C, et al. Memristor-based analog computation and neural network classification with a dot product engine. Advanced Materials. 2018;30(9):1705914. Yu S. Neuro-inspired computing with emerging nonvolatile memories. Proceedings of the IEEE. 2018;106(2):260–285. Pei J, Deng L, Song S, et al. Towards artificial general intelligence with hybrid Tianjic chip architecture. Nature. 2019;572:106–111. Yao P, Wu H, Gao B, et al. Fully hardware-implemented memristor convolutional neural network. Nature. 2020;577:641–646. Roy K, Jaiswal A and Panda P. Towards spike-based machine intelligence with neuromorphic computing. Nature. 2019 11;575:607–617. Sherstinsky A. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. CoRR. 2018;abs/1808.03314. Yu S. Resistive random access memory (RRAM). Synthesis Lectures on Emerging Engineering Technologies. 2016;2:1–79. Kim KH, Gaba S, Wheeler D, et al. A functional hybrid memristor crossbararray/CMOS system for data storage and neuromorphic applications. Nano Letters. 2012;12(1):389–395. Yang JJ, Zhang MX, Pickett M, et al. Engineering nonlinearity into memristors for passive crossbar applications. Applied Physics Letters. 2012 03;100. Park S, Kim H, Choo M, et al. RRAM-based synapse for neuromorphic system with pattern recognition function. In: International Electron Devices Meeting; 2012. Khvalkovskiy A, Apalkov D, Watts S, et al. Basic principles of STT-MRAM cell operation in memory arrays. Journal of Physics D: Applied Physics. 2013;46(7):1–35 Apalkov D, Khvalkovskiy A, Watts S, et al. Spin-transfer torque magnetic random access memory (STT-MRAM). J Emerg Technol Comput Syst. 2013;9(2). Bi G and Poo M. Synaptic modifications in cultured hippocampal neurons: Dependence on spike timing, Synaptic Strength, and Postsynaptic Cell Type. J Neuroscience. 1998;18(24):10467-10472. Hu M, Strachan JP, Li Z, et al. Dot-product engine for neuromorphic computing: Programming 1T1M crossbar to accelerate matrix-vector multiplication. In: Annual Design Automation Conference; 2016. Kuzum D, Jeyasingh R, Lee B, et al. Nanoelectronic programmable synapses based on phase change materials for brain-inspired computing. Nano Letters. 2011 06;12:2179–86.
References [91]
[92]
[93]
[94]
[95] [96]
[97] [98]
[99]
[100] [101]
[102] [103]
[104]
[105]
[106]
223
Indiveri G, Chicca E and Douglas R. A VLSI array of low-power spiking neurons and bistable synapses with spike-timing dependent plasticity. IEEE Transactions on Neural Networks. 2006;17(1):211–221. Chen P, Lin B, Wang I, et al. Mitigating effects of non-ideal synaptic device characteristics for on-chip learning. In: IEEE/ACM International Conference on Computer-Aided Design (ICCAD); 2015. Wang IT, Chang CC, Chiu LW, et al. 3D Ta/TaOx /TiO2 /Ti synaptic array and linearity tuning of weight update for hardware neural network applications. Nanotechnology. 2016;27:365204. Angizi S, He Z, Parveen F and Fan D. IMCE: Energy-efficient bit-wise inmemory convolution engine for deep neural network. In: Asia and South Pacific Design Automation Conference (ASP-DAC); 2018. Patterson D, Anderson T, Cardwell N, et al. A case for intelligent RAM. IEEE Micro. 1997. Pudukotai Dinakarrao SM, Lin J, Zhu S, et al. A scalable network-onchip microprocessor with 2.5D integrated memory and accelerator. IEEE Transactions on Circuits and Systems I: Regular Papers. 2017. Yong-Bin K and Chen T. Assessing merged DRAM/Logic technology. Integr VLSI J. 1999. Deng Q, Zhang Y, Zhang M and Yang J. LAcc: Exploiting lookup table-based fast and accurate vector multiplication in DRAM-based CNN accelerator. In: ACM/IEEE Design Automation Conference (DAC); 2019. Courbariaux M, Bengio Y and David J-P. BinaryConnect: Training deep neural networks with binary weights during propagations. In: Proceedings of the 28th International Conference on Neural Information Processing Systems; 2015. Courbariaux M and BengioY. BinaryNet: Training deep neural networks with weights and activations constrained to +1 or -1. ArXiv. 2016. Rastegari M, Ordonez V, Redmon J and Farhadi A. XNOR-Net: ImageNet classification using binary convolutional neural networks. In: Computer Vision – ECCV; 2016. Li F and Liu B. Ternary weight networks. CoRR. 2016. Wess M, Dinakarrao SMP and Jantsch A. Weighted quantizationregularization in DNNs for weight memory minimization toward HW implementation. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. 2018. Li S, Niu D, Malladi KT et al. DRISA: A DRAM-based reconfigurable in-situ accelerator. In: IEEE/ACM International Symposium on Microarchitecture; 2017. Deng Q, Jiang L, Zhang Y, Zhang M and Yang J. DrAcc: A DRAMbased accelerator for accurate CNN inference. In: ACM/ESDA/IEEE Design Automation Conference (DAC);2018. Yin S, Jiang Z, Seo J and Seok M. XNOR-SRAM: In-Memory computing SRAM macro for binary/ternary deep neural networks. IEEE Journal of SolidState Circuits. 2020.
224 [107]
[108]
[109]
[110]
[111]
[112]
[113]
[114]
[115]
[116]
[117]
[118]
[119]
ReRAM-based machine learning Sun X, Yin S, Peng X, Liu R, Seo J and Yu S. XNOR-RRAM: A scalable and parallel resistive synaptic architecture for binary neural networks. In: Design, Automation Test in Europe Conference Exhibition (DATE); 2018. Tang T, Xia L, Li B, WangY andYang H. Binary convolutional neural network on RRAM. In: Asia and South Pacific Design Automation Conference (ASPDAC); 2017. Pan Y, Ouyang P, Zhao Y, et al. A multilevel cell STT-MRAM-based computing in-memory accelerator for binary convolutional neural network. IEEE Transactions on Magnetics. 2018. Angizi S, He Z, Awad A and Fan D. MRIMA: An MRAM-based in-memory accelerator. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. 2020. Taherinejad N, Manoj PDS and Jantsch A. Memristors’ potential for multi-bit storage and pattern learning. IEEE European Modelling Symposium (EMS). 2015. Chi P, Li S, Xu C, et al. PRIME: A novel processing-in-memory architecture for neural network computation in ReRAM-based main memory. In: ACM/IEEE International Symposium on Computer Architecture (ISCA); 2016. Simon WA, Qureshi YM, Levisse A, et al. Blade: A bitline accelerator for devices on the edge. In: Proceedings of the 2019 on Great Lakes Symposium on VLSI; 2019. pp. 207–212. Yang J, Kong Y, Wang Z, et al. 24.4 sandwich-RAM: An energy-efficient in-memory BWN architecture with pulse-width modulation. In: 2019 IEEE International Solid-State Circuits Conference-(ISSCC). IEEE; 2019. pp. 394–396. Bankman D, Yang L, Moons B, et al. An always-on 3.8uJ/86% CIFAR-10 mixed-signal binary CNN processor with all memory on chip in 28-nm CMOS. IEEE Journal of Solid-State Circuits. 2018;54(1):158–172. Si X, Chen JJ, Tu YN, et al. 24.5 A twin-8T SRAM computation-inmemory macro for multiple-bit CNN-based machine learning. In: 2019 IEEE International Solid-State Circuits Conference-(ISSCC). IEEE; 2019. pp. 396–398. Si X, Tu YN, Huanq WH, et al. 15.5 A 28nm 64Kb 6T SRAM computingin-memory macro with 8b MAC operation for AI edge chips. In: 2020 IEEE International Solid-State Circuits Conference (ISSCC). IEEE; 2020. pp. 246–248. Dong Q, Sinangil ME, Erbagci B, et al. 15.3 A 351TOPS/W and 372.4 GOPS compute-in-memory SRAM macro in 7nm FinFET CMOS for machinelearning applications. In: 2020 IEEE International Solid-State Circuits Conference-(ISSCC). IEEE; 2020. pp. 242–244. Zhang Z, Chen JJ, Si X, et al. A 55nm 1-to-8 bit configurable 6T SRAM based computing-in-memory unit-macro for CNN-based AI edge processors. In: 2019 IEEE Asian Solid-State Circuits Conference (A-SSCC). IEEE; 2019. pp. 217–218.
References [120]
[121]
[122]
[123]
[124]
[125]
[126]
[127]
[128]
[129]
[130]
[131]
[132]
[133]
225
Xue CX, Chen WH, Liu JS, et al. Embedded 1-Mb ReRAM-based computing-in-memory macro with multibit input and weight for CNNbased AI edge processors. IEEE Journal of Solid-State Circuits. 2019;55(1): 203–215. Mochida R, Kouno K, Hayata Y, et al. A 4M synapses integrated analog ReRAM based 66.5 TOPS/W neural-network processor with cell current controlled writing and flexible network architecture. In: 2018 IEEE Symposium on VLSI Technology. IEEE; 2018. pp. 175–176. Eckert C, Wang X, Wang J, et al. Neural cache: Bit-serial in-cache acceleration of deep neural networks. In: International Symposium on Computer Architecture (ISCA); 2018. Seshadri V, Lee D, Mullins T, et al. Ambit: In-memory accelerator for bulk bitwise operations using commodity DRAM technology. IEEE/ACM International Symposium on Microarchitecture (MICRO). 2017. Li S, Glova AO, Hu X, et al. SCOPE: A stochastic computing engine for DRAM-based in-situ accelerator. In: IEEE/ACM International Symposium on Microarchitecture (MICRO); 2018. Sutradhar PR, Connolly M, Bavikadi S, et al. pPIM: A programmable processor-in-memory architecture with precision-scaling for deep learning. IEEE Computer Architecture Letters. 2020;19(2):118–121. Ali M, Jaiswal A, Kodge S, Agrawal A, Chakraborty I and Roy K. IMAC: Inmemory multi-bit multiplication and accumulation in 6T SRAM array. IEEE Transactions on Circuits and Systems I: Regular Papers. 2020. Shafiee A, Nag A, Muralimanohar N, et al. ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars. In: ACM/IEEE International Symposium on Computer Architecture (ISCA); 2016. Song L, Qian X, Li H and Chen Y. PipeLayer: A pipelined ReRAM-based accelerator for deep learning. In: IEEE Int. Symp. on High Performance Computer Architecture (HPCA); 2017. Qiao X, Cao X, Yang H, Song L and Li H. AtomLayer: A universal ReRAMbased CNN accelerator with atomic layer computation. In: ACM/ESDA/IEEE Design Automation Conference (DAC); 2018. Patil AD, Hua H, Gonugondla S, Kang M and Shanbhag NR. An MRAM-based deep in-memory architecture for deep neural networks. In: IEEE International Symposium on Circuits and Systems; 2019. Fan D, Angizi S. Energy efficient in-memory binary deep neural network accelerator with dual-mode SOT-MRAM. In: IEEE International Conference on Computer Design (ICCD); 2017. Angizi S, He Z, Rakin AS and Fan D. CMP-PIM: An energy-efficient comparator-based processing-in-memory neural network accelerator. In: ACM/ESDA/IEEE Design Automation Conference (DAC); 2018. Yu H, Ni L and Huang H. In: Vaidyanathan S, Volos C, editors. Distributed in-memory computing on binary memristor-crossbar for machine
226
[134] [135] [136] [137]
[138] [139]
[140]
[141]
[142]
[143]
[144]
[145]
[146]
[147]
[148]
ReRAM-based machine learning learning. Springer; 2017. pp. 275–304. Available from: https://link. springer.com/chapter/10.1007/978-3-319-51724-7_12. Strukov DB and Williams RS. Exponential ionic drift: Fast switching and low volatility o thin-film memristors. Applied Physics A. 2009;94(3):515–519. Joglekar YN and Wolf SJ. The elusive memristor: Properties of basic electrical circuits. European Journal of Physics. 2009;30(4):661. Biolek Z, Biolek D and Biolkova V. SPICE model of memristor with nonlinear dopant drift. Radioengineering. 2009;18(2):210–214. Wang Z, Joshi S, Savel ev SE, et al. Memristors with diffusive dynamics as synaptic emulators for neuromorphic computing. Nature Materials. 2017;16(1):101–108. Midya R, Wang Z, Zhang J, et al. Anatomy of Ag/Hafnia-based selectors with 1010 nonlinearity. Advanced Materials. 2017;29(12). Govoreanu B, Kar GS, Chen YY, et al. 10 × 10 nm2 Hf/HfOx crossbar resistive RAM with excellent performance, reliability and low-energy operation. In: International Electron Devices Meeting (IEDM); 2011. pp. 31.6.1–31.6.4. Kang J, Gao B, Chen B, et al. 3D RRAM: Design and optimization. In: IEEE Conference on International Solid-State and Integrated Circuit Technology (ICSICT); 2014. pp. 1–4. Fan D, Sharad M and Roy K. Design and synthesis of ultralow energy spin-memristor threshold logic. IEEE Transactions on Nanotechnology. 2014;13(3):574–583. Gu P, Li B, Tang T, et al. Technological exploration of RRAM crossbar array for matrix-vector multiplication. In: Asia and South Pacific Design Automation Conference (ASP-DAC). IEEE; 2015. pp. 106–111. Srimani T, Manna B, Mukhopadhyay AK, et al. Energy efficient and high performance current-mode neural network circuit using memristors and digitally assisted analog CMOS neurons. arXiv preprint arXiv:151109085. 2015. Wang Y, Yu H and Zhang W. Nonvolatile CBRAM-crossbar-based 3-Dintegrated hybrid memory for data retention. IEEE Transactions on Very Large Scale Integration (VLSI) Systems. 2014;22(5):957–970. Hu M, Strachan JP, Li Z, et al. Dot-product engine for neuromorphic computing: Programming 1T1M crossbar to accelerate matrix-vector multiplication. In: Design Automation Conference (DAC); 2016. pp. 19.1–19.6. Xia L, Tang T, Huangfu W, et al. Switched by input: Power efficient structure for RRAM-based convolutional neural network. In: Proceedings of the 53rd Annual Design Automation Conference. ACM; 2016. pp. 125.1–125.6. Ni L, Wang Y, Yu H, et al. An energy-efficient matrix multiplication accelerator by distributed in-memory computing on binary RRAM crossbar. In: Asia and South Pacific Design Automation Conference (ASP-DAC). IEEE; 2016. pp. 280–285. Ni L, Huang H and Yu H. On-line machine learning accelerator on digital RRAM-crossbar. In: International Symposium on Circuits and Systems (ISCAS). IEEE; 2016. pp. 113–116.
References [149]
[150]
[151]
[152] [153] [154] [155] [156] [157]
[158]
[159] [160]
[161]
[162]
[163]
[164]
227
Tang T, Xia L, Li B, et al. Binary convolutional neural network on RRAM. In: Asia and South Pacific Design Automation Conference (ASP-DAC); 2017. pp. 782–787. Yan B, Mahmoud AM,Yang JJ, et al. A neuromorphic ASIC design using oneselector-one-memristor crossbar. In: International Symposium on Circuits and Systems (ISCAS); 2016. pp. 1390–1393. Jackson TC, Sharma AA, Bain JA, et al. Oscillatory neural networks based on TMO nano-oscillators and multi-level RRAM cells. IEEE Journal on Emerging and Selected Topics in Circuits and Systems. 2015;5(2): 230–241. Sharma AA, Li Y, Skowronski M, et al. High-frequency TaOx-based compact oscillators. IEEE Transactions on Electron Devices. 2015;62(11):3857–3862. Maffezzoni P, Daniel L, Shukla N, et al. Modelling hysteresis in vanadium dioxide oscillators. Electronics Letters. 2015;51(11):819–820. Chua L. Resistance switching memories are memristors. Applied Physics A. 2011;102(4):765–783. Pershin YV, Fontaine SL and Ventra MD. Memristive model of Amoeba learning. Physical Review E. 2009;80:1–6. Borghetti J, Snider GS, Kuekes PJ, et al. Memristive switches enable stateful logic operations via material implications. Nature. 2010;464:873–876. Pershin YV and Di Ventra M. Neuromorphic, digital, and quantum computation with memory circuit elements. Proceedings of the IEEE. 2012;100(6):2071–2080. Hongal V, Kotikalapudi R and Choi M. Design, test, and repair of MLUT (memristor look-up table) based asynchronous nanowire reconfigurable crossbar architecture. IEEE Journal on Emerging and Selected Topics in Circuits and Systems,. 2014;4(4):427–437. Levy Y, Bruck J, Cassuto Y, et al. Logic operations in memory using a memristive Akers array. Microelectronics Journal. 2014;45:873–876. Ho Y, Huang GM and Li P. Dynamical properties and design analysis for nonvolatile memristor memories. IEEE Transactions on Circuits and Systems I: Regular Papers,. 2011;58(4):724–736. Mohammad B, Homouz D and Elgabra H. Robust hybrid memristor-CMOS memory: Modeling and design. IEEE Transactions on Very Large Scale Integration (VLSI) Systems. 2013;21(11):2069–2079. Baghel VS and Akashe S. Low power memristor based 7T SRAM using MTCMOS technique. In: 2015 Fifth International Conference on Advanced Computing Communication Technologies (ACCT); 2015. pp. 222–226. International Technology Roadmap for Semiconductors—System Drivers; 2011. Available from: https://www.semiconductors.org/wp-content/uploads/ 2018/08/2011SysDrivers.pdf. Niu D, Chen Y and Xie Y. Low-power dual-element memristor based memory design. In: 2010 ACM/IEEE International Symposium on Low-Power Electronics and Design (ISLPED); 2010. pp. 25–30.
228 [165]
[166]
[167]
[168]
[169]
[170]
[171]
[172]
[173]
[174] [175] [176] [177] [178] [179] [180] [181]
ReRAM-based machine learning Kim H, Sah MP, Yang C, et al. Memristor-based multilevel memory. In: 2010 12th International Workshop on Cellular Nanoscale Networks and Their Applications (CNNA); 2010. pp. 1–6. Zangeneh M and Joshi A. Design and optimization of nonvolatile multibit 1T1R resistive RAM. IEEE Transactions on Very Large Scale Integration (VLSI) Systems. 2014;22(8):1815–1828. Rothenbuhler A, Tran T, Smith EHB, et al. Reconfigurable threshold logic gates using memristive devices. Journal of Low Power Electronics and Applications. 2013;3(2):174. Biolek D, Di Ventra M and Pershin YV. Reliable SPICE simulations of memristors, memcapacitors and meminductors. ArXiv e-prints. 2013;22(4): 945–968. Taherinejad N, Sai Manoj PD and Jantsch A. Memristors’ potential for multibit storage and pattern learning. Proceedings of the IEEE European Modelling Symposium (EMS) Conference. 2015 Oct. Kvatinsky S, Wald N, Satat G, et al. MRL—Memristor ratioed logic. In: IEEE Int. W. on Cellular Nanoscale Networks and Their Applications; 2012. Zhang Y, Shen Y, Wang X, et al. A novel design for memristor-based logic switch and crossbar circuits. IEEE Tran on Circuits and Systems I. 2015;62(5):1402–1411. Vourkas I, Batsos A and Sirakoulis GC. SPICE modeling of nonlinear memristive behavior. Int J of Circuit Theory and Applications. 2015;43(5): 553–565. Vourkas I and Siakoulis GC. Memristor-based nanoelectronic computing architectures. Vol. 19. Emergence, Complexity and Computation. Springer; 2015. pp. 9–26. Available from: https://www.springer.com/gp/book/ 9783319226460. Haykin SS. Neural Networks and Learning Machines. Vol. 3. Pearson Education, Upper Saddle River; 2009. Wold S, Esbensen K and Geladi P. Principal component analysis. Chemometrics and Intelligent Laboratory Systems. 1987;2(1-3):37–52. Suykens JA and Vandewalle J. Least squares support vector machine classifiers. Neural Processing Letters. 1999;9(3):293–300. LeCun YA, Bottou L, Orr GB, et al. Efficient backprop. Neural Networks: Tricks of the Trade. Springer; 2012. pp. 9–48. Werbos PJ. Backpropagation through time: What it does and how to do it. Proceedings of the IEEE. 1990;78(10):1550–1560. Higham NJ. Cholesky factorization. Wiley Interdisciplinary Reviews: Computational Statistics. 2009;1(2):251–254. Krishnamoorthy A and Menon D. Matrix inversion using Cholesky decomposition. arXiv preprint arXiv:11114144. 2011. Cong J and Xiao B. Minimizing computation in convolutional neural networks. In: International Conference on Artificial Neural Networks (ICANN). Springer; 2014. pp. 281–290.
References [182] [183] [184] [185] [186]
[187]
[188]
[189]
[190]
[191]
[192]
[193] [194] [195]
[196]
[197] [198] [199]
229
Michalski RS, Carbonell JG and Mitchell TM. Machine learning: An artificial intelligence approach. Springer Science & Business Media; 2013. Zhang L and Suganthan P. A comprehensive evaluation of random vector functional link networks. Information Sciences. 2016;367:1094–1105. Bache K and Lichman M. UCI Machine Learning Repository; 2013. Available from: http://archive.ics.uci.edu/ml. LeCun Y, Bengio Y and Hinton G. Deep learning. Nature. 2015;521(7553): 436–444. Lawrence S, Giles CL, Tsoi AC, et al. Face recognition: A convolutional neural-network approach. IEEE Transactions on Neural Networks. 1997;8(1):98–113. Krizhevsky A, Sutskever I and Hinton GE. Imagenet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems (NIPS); 2012. pp. 1097–1105. Courbariaux M, Bengio Y and David JP. Binaryconnect: Training deep neural networks with binary weights during propagations. In: Advances in Neural Information Processing Systems (NIPS); 2015. pp. 3123–3131. Liu Z, Li Y, Ren F, et al. A binary convolutional encoder-decoder network for real-time natural scene text processing. arXiv preprint arXiv:161203630. 2016. Ioffe S and Szegedy C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:150203167. 2015. Qiu J, Wang J, Yao S, et al. Going deeper with embedded FPGA platform for convolutional neural network. In: International Symposium on FieldProgrammable Gate Arrays (FPGA); 2016. pp. 26–35. Plaut DC. Experiments on Learning by Back Propagation. Carnegie Mellon University; 1986. Available from: http://www.cnbc.cmu.edu/∼plaut/ papers/pdf/PlautNowlanHinton86TR.backprop.pdf. Hubara I, Soudry D and Yaniv RE. Binarized neural networks. arXiv preprint arXiv:160202505. 2016. Davis A and Arel I. Low-rank approximations for conditional feedforward computation in deep neural networks. arXiv preprint arXiv:13124461. 2013. Nakkiran P, Alvarez R, Prabhavalkar R, et al. Compressing deep neural networks using a rank-constrained topology. In: INTERSPEECH; 2015. pp. 1473–1477. Novikov A, Podoprikhin D, Osokin A, et al. Tensorizing neural networks. In: Advances in Neural Information Processing Systems (NIPS); 2015. pp. 442–450. Oseledets IV. Tensor-train decomposition. SIAM Journal on Scientific Computing. 2011;33(5):2295–2317. Cichocki A. Era of big data processing: A new approach via tensor networks and tensor decompositions. arXiv preprint arXiv:14032048. 2014. Hagan MT, Demuth HB, Beale MH, et al. Neural Network Design. Vol. 20. PWS publishing company, Boston; 1996.
230 [200]
[201]
[202]
[203]
[204]
[205]
[206]
[207]
[208] [209]
[210]
[211]
[212]
[213] [214]
[215]
ReRAM-based machine learning Bengio Y, Lamblin P, Popovici D, et al. Greedy Layer-Wise Training of Deep Networks. In: Advances in Neural Information Processing Systems; 2007. pp. 153–160. Tang J, Deng C and Huang GB. Extreme learning machine for multilayer perceptron. IEEE Transactions on Neural Networks and Learning Systems. 2016;27(4):809–821. Erhan D, Manzagol PA, Bengio Y, et al. The difficulty of training deep architectures and the effect of unsupervised pre-training. In: AISTATS. vol. 5; 2009. pp. 153–160. Holtz S, Rohwedder T and Schneider R. The alternating linear scheme for tensor optimization in the tensor train format. SIAM Journal on Scientific Computing. 2012;34(2):A683–A713. Oseledets IV and Dolgov S. Solution of linear systems and matrix inversion in the TT-format. SIAM Journal on Scientific Computing. 2012;34(5):A2718– A2739. Wright J, Yang AY, Ganesh A, et al. Robust face recognition via sparse representation. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2009;31(2):210–227. Lee H, Chen P, Wu T, et al. Low power and high speed bipolar switching with a thin reactive Ti buffer layer in robust HfO2 based RRAM. In: International Electron Devices Meeting (IEDM). IEEE; 2008. pp. 1–4. Singh PN, Kumar A, Debnath C, et al. 20mW, 125Msps, 10bit pipelined ADC in 65nm standard digital CMOS process. In: Custom Integrated Circuits Conference (CICC). IEEE; 2007. pp. 189–192. Tan T and Sun Z. CASIA-FingerprintV5; 2010. Available from: http: //biometrics.idealtest.org/. Chiarulli DM, Jennings B, Fang Y, et al. A computational primitive for convolution based on coupled oscillator arrays. In: Computer Society Annual Symposium on VLSI (ISVLSI). IEEE; 2015. pp. 125–130. Yogendra K, Fan D, Shim Y, et al. Computing with coupled Spin Torque Nano Oscillators. In: Asia and South Pacific Design Automation Conference (ASP-DAC). IEEE; 2016. pp. 312–317. Topaloglu RO. More than Moore technologies for next generation computer design. Springer; 2015. Available from: https://www.springer.com/gp/ book/9781493921621. Xu D, Yu N, Huang H, et al. Q-Learning based voltage-swing tuning and compensation for 2.5D memory-logic integration. IEEE Design and Test. 2018;35(2):91–99. Xu D, Yu N, Manoj PDS, et al. A 2.5-D memory-logic integration with datapattern-aware memory controller. IEEE Design Test. 2015;32(4):1–10. Manoj PDS, Yu H, Huang H, et al. A Q-learning based self-adaptive I/O communication for 2.5D integrated many-core microprocessor and memory. IEEE Trans on Computers. 2016;65(4):1185–1196. Manoj PDS, Yu H and Wang K. 3D many-core microprocessor power management by space-time multiplexing based demand-supply matching. IEEE Trans on Computers. 2015;64(11):3022–3036.
References [216]
[217]
[218] [219]
[220]
[221]
[222]
[223]
[224]
[225]
[226]
[227]
[228]
[229]
[230]
231
Manoj PDS, Yu H, Shang Y, et al. Reliable 3-D clock-tree synthesis considering nonlinear capacitive TSV model with electrical-thermal-mechanical coupling. IEEE Trans on Computer-Aided Design of Integrated Circuits and Systems. 2013;32(11):1734–1747. LiauwYY, Zhang Z, Kim W, et al. Nonvolatile 3D-FPGA with monolithically stacked RRAM-based configuration memory. In: International Solid-State Circuits Conference (ISSCC). IEEE; 2012. pp. 406–408. Poremba M, Mittal S, Li D, et al. DESTINY: A tool for modeling emerging 3D NVM and eDRAM caches. In: IEEE DATE; 2015. pp. 1543–1546. ChenYC, Wang W, Li H, et al. Non-volatile 3D stacking RRAM-based FPGA. In: International Conference on Field Programmable Logic and Applications (FPL). IEEE; 2012. pp. 367–372. TaheriNejad N, Manoj PDS, Rathmair M, et al. Fully digital write-in scheme for multi-bit memristive storage. In: Conference on Electrical Engineering, Computing Science and Automatic Control; 2016. TaheriNejad N, Manoj PDS and Jantsch A. Memristor’s potential for multibit storage and pattern learning. In: European Modeling Symposium on Mathematical Modeling and Computer Simulation; 2015. Ni L, Liu Z, Yu H, et al. An energy-efficient digital ReRAM-crossbar based CNN with bitwise parallelism. IEEE Journal on Exploratory Solid-State Computational Devices and Circuits. 2017;3(1):1–10. Yu S, Li Z, Chen PY, et al. Binary neural network with 16 Mb RRAM macro chip for classification and online training. In: International Electron Devices Meeting; 2016. pp. 16.2.1–16.2.4. Rastegari M, Ordonez V, Redmon J, et al. XNOR-net: ImageNet classification using binary convolutional neural networks. In: European Conference on Computer Vision (ECCV); 2016. pp. 525–542. Ni L, Huang H and Yu H. A memristor network with coupled oscillator and crossbar towards L2-norm based machine learning. In: International Symposium on Nanoscale Architectures (NANOARCH). IEEE; 2016. pp. 179–184. Huang GB, Ramesh M, Berg T, et al. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. University of Massachusetts, Amherst; 2007. Vedaldi A and Lenc K. MatConvNet: Convolutional neural networks for Matlab. In: International Conference on Multimedia. ACM; 2015. pp. 689–692. Goll B and Zimmermann H. A 65nm CMOS comparator with modified latch to achieve 7 GHz/1.3 mW at 1.2 V and 700 MHz/47 μW at 0.6 V. In: International Solid-State Circuits Conference (ISSCC). IEEE; 2009. pp. 328–329. Zhang C, Li P, Sun G, et al. Optimizing FPGA-based accelerator design for deep convolutional neural networks. In: International Symposium on FieldProgrammable Gate Arrays (FPGA). ACM; 2015. pp. 161–170. Chen YH, Krishna T, Emer J, et al. Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. In:
232
[231]
[232] [233] [234]
[235]
[236]
[237]
[238] [239]
[240] [241] [242]
[243]
[244]
[245]
ReRAM-based machine learning International Solid-State Circuits Conference (ISSCC). IEEE; 2016. pp. 262–263. Hinton GE and Salakhutdinov RR. Replicated softmax: An undirected topic model. In: Advances in Neural Information Processing Systems (NIPS); 2009. pp. 1607–1614. LeCun Y, Cortes C and Burges CJ. MNIST handwritten digit database; 2010. Available from: http://yann.lecun.com/exdb/mnist. Krizhevsky A, Nair V and Hinton G. The CIFAR-10 dataset; 2014. Available from: http://www.cs.toronto.edu/∼kriz/cifar.html. Chen YH, Krishna T, Emer JS, et al. Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. IEEE Journal of Solid-State Circuits. 2017;52(1):127–138. Huang H, Ni L, Wang Y, et al. A 3D multi-layer CMOS-RRAM accelerator for neural network. In: International Conference on 3D System Integration (3DIC). IEEE; 2016. Chen PY, Kadetotad D, Xu Z, et al. Technology-design co-optimization of resistive cross-point array for accelerating learning algorithms on chip. In: Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE; 2015. pp. 854–859. Ni L, Huang H, Liu Z, et al. Distributed in-memory computing on binary RRAM crossbar. ACM Journal on Emerging Technologies in Computing Systems (JETC). 2017;13(3):36.1–36.18. Franzon P, Rotenberg E, Tuck J, et al. Computing in 3D. In: Custom Integrated Circuits Conference (CICC). IEEE; 2015. pp. 1–6. Kim DH, Athikulwongse K and Lim SK. Study of through-silicon-via impact on the 3-D stacked IC layout. IEEE Transactions on Very Large Scale Integration (VLSI) Systems. 2013;21(5):862–874. Krizhevsky A and Hinton G. Learning multiple layers of features from tiny images. University of Toronto; 2009. GPU specs. Available from: http://www.nvidia.com/object/workstationsolutions.html. Chen K, Li S, Muralimanohar N, et al. CACTI-3DD: Architecture-level modeling for 3D die-stacked DRAM main memory. In: Proceedings of the Conference on Design, Automation and Test in Europe. IEEE; 2012. pp. 33–38. Xue J, Li J and GongY. Restructuring of deep neural network acoustic models with singular value decomposition. In: INTERSPEECH; 2013. pp. 2365– 2369. CPU Xeon-X5690specs. Accessed: 2017-03-30. Available from: http://ark. intel.com/products/52576/Intel-Xeon-Processor-X5690-12M-Cache-3_46GHz-6_40-GTs-Intel-QPI. Han S, Mao H and Dally WJ. Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding. arXiv preprint arXiv:151000149. 2015.
References [246]
[247] [248]
[249]
[250]
[251]
[252]
[253] [254]
[255]
[256]
[257]
[258]
[259]
233
He K, Zhang X, Ren S, et al. Deep residual learning for image recognition. In: IEEE Conference on Computer Vision and Pattern Recognition; 2016. pp. 770–778. He K, Zhang X and Ren S. Identity mappings in deep residual networks. In: European Conference on Computer Vision; 2016. pp. 630–645. Guo K, Sui L, Qiu J, et al. From model to FPGA: Software-hardware codesign for efficient neural network acceleration. In: IEEE Symposium on Hot Chips; 2016. pp. 1–27. Wang Y, Li X, Yu H, et al. Optimizing Boolean embedding matrix for compressive sensing in RRAM crossbar. In: IEEE/ACM International Symposium on Low Power Electronics and Design. IEEE; 2015. pp. 13–18. Chen T, Du Z, Sun N, et al. Diannao: A small-footprint high-throughput accelerator for ubiquitous machine-learning. ACM SIGPLAN Notices. 2014;49(4):269–284. Wen W, Wu C, Wang Y, et al. Learning structured sparsity in deep neural networks. In: Advances in Neural Information Processing Systems; 2016. pp. 2074–2082. Hashemi S, Anthony N, Tann H, et al. Understanding the impact of precision quantization on the accuracy and energy of neural networks. In: Design, Automation and Test in Europe; 2017. pp. 1474–1479. Dettmers T. 8-bit approximations for parallelism in deep learning. arXiv preprint arXiv:151104561. 2015. Hubara I, Courbariaux M, Soudry D, El-Yaniv R and Bengio Y. Quantized neural networks: Training neural networks with low precision weights and activations. Journal of Machine Learning Research. 2017;18(1): 6869–6898. Jacob B, Kligys S, Chen B, et al. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2018. pp. 2704– 2713. Zhou S, Wu Y, Ni Z, et al. DoReFa-Net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:160606160. 2016. Bengio Y, Léonard N and Courville A. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:13083432. 2013. Chi P, Li S, Xu C, et al. PRIME: A novel processing-in-memory architecture for neural network computation in ReRAM-based main memory. In: ACM/SIGARCH Computer Architecture News. vol. 44; 2016. pp. 27–39. Albericio J, Judd P, Hetherington T, et al. Cnvlutin: Ineffectual-neuron-free deep neural network computing. In: ACM/SIGARCH Computer Architecture News. vol. 44; 2016. pp. 1–13.
234 [260]
[261]
[262]
[263] [264]
[265]
[266] [267]
[268]
[269] [270] [271]
[272] [273]
[274]
[275]
ReRAM-based machine learning Fan D and Angizi S. Energy-efficient in-memory binary deep neural network accelerator with dual-mode SOT-MRAM. In: IEEE International Conference on Computer Design. IEEE; 2017. pp. 609–612. Wang Y, Ni L, Chang CH, et al. DW-AES: A domain-wall nanowire-based AES for high throughput and energy-efficient data encryption in nonvolatile memory. IEEE Transactions on Information Forensics and Security. 2016;11(11):2426–2440. Sharma A, Jackson T, Schulaker M, et al. High performance, integrated 1T1R oxide-based oscillator: Stack engineering for low-power operation in neural network applications. In: IEEE Symposium on VLSI Technology; 2015. pp. T186–T187. Chang SC, Kani N, Manipatruni S, et al. Scaling limits on all-spin logic. IEEE Transactions on Magnetics. 2016;52(7):1–4. Huang H, Ni L, Wang K, et al. A highly parallel and energy efficient threedimensional multilayer CMOS-RRAM accelerator for tensorized neural network. IEEE Transactions on Nanotechnology. 2018;17(4):645–656. Deng J, Dong W, Socher R, et al. Imagenet: A large-scale hierarchical image database. In: IEEE Conference on Computer Vision and Pattern Recognition; 2009. pp. 248–255. Zhu C, Han S, Mao H, et al. Trained ternary quantization. arXiv preprint arXiv:161201064. 2016. Nair V and Hinton GE. Rectified linear units improve restricted Boltzmann machines. In: International Conference on Machine Learning; 2010. pp. 807–814. Xu C, Niu D, Muralimanohar N, et al. Overcoming the challenges of crossbar resistive memory architectures. In: IEEE Symposium on High Performance Computer Architecture; 2015. pp. 476–488. Kinga D and Adam JB. A method for stochastic optimization. In: International Conference on Learning Representations. vol. 5; 2015. Li C, Hu M, Li Y, et al. Analogue signal and image processing with large memristor crossbars. Nature Electronics. 2018;1(1):52. Lee PC, Lin JY and Hsieh CC. A 0.4 V 1.94 fJ/conversion-step 10 bit 750 kS/s SAR ADC with input-range-adaptive switching. IEEE Transactions on Circuits and Systems I: Regular Papers. 2016;63(12):2149–2157. Stathopoulos S, Khiat A, Trapatseli M, et al. Multibit memory operation of metal-oxide bi-layer memristors. Scientific reports. 2017;7(1):17532. Ma Y, Cao Y, Vrudhula S, et al. Performance modeling for CNN inference accelerators on FPGA. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. 2019. Kaplan R, Yavits L and Ginosar R. PRINS: Processing-in-storage acceleration of machine learning. IEEE Transactions on Nanotechnology. 2018;17(5):889–896. Song L, Qian X, Li H, et al. Pipelayer: A pipelined ReRAM-based accelerator for deep learning. In: IEEE International Symposium on High Performance Computer Architecture; 2017. pp. 541–552.
References [276] [277]
[278] [279]
[280]
[281]
[282]
[283]
[284] [285]
[286]
[287]
[288] [289]
[290]
[291]
235
Mellempudi N, Kundu A, Mudigere D, et al. Ternary neural networks with fine-grained quantization. arXiv preprint arXiv:170501462. 2017. Zhou A, Yao A, Guo Y, et al. Incremental network quantization: Towards lossless CNNS with low-precision weights. arXiv preprint arXiv:170203044. 2017. Li F, Zhang B and Liu B. Ternary weight networks. arXiv preprint arXiv:160504711. 2016. Dong X, Xu C, Jouppi N, et al. NVSim: A circuit-level performance, energy, and area model for emerging non-volatile memory. In: Emerging Memory Technologies; 2014. pp. 15–50. Ren F and Markovic D. A configurable 12-to-237KS/s 12.8 mW sparseapproximation engine for mobile ExG data aggregation. In: Proc. 2015 IEEE Int. Solid-State Circuits Conf., San Francisco, CA; 2015. pp. 1–3. Zhang Z, Jung T, Makeig S and Rao BD. Compressed sensing of EEG for wireless telemonitoring with low energy consumption and inexpensive hardware. IEEE Trans Biomed Eng. 2013;60(1):221–224. Dixon AMR, Allstot EG, Gangopadhyay D and Allstot DJ. Compressed sensing system considerations for ECG and EMG wireless biosensors. IEEE Trans Biomed Circuits Syst. 2012;6(2):156–166. Suo Y, Zhang J, Xiong T, Chin PS, Etienne-Cummings R and Tran TD. Energy-efficient multi-mode compressed sensing system for implantable neural recordings. IEEE Trans Biomed Circuits Syst. 2014;8(5): 648–659. Donoho DL. Compressed sensing. IEEE Trans Inf Theory. 2006;52(4): 1289–1306. Pant JK and Krishnan S. Compressive sensing of electrocardiogram signals by promoting sparsity on the second-order difference and by using dictionary learning. IEEE Trans Biomed Circuits Syst. 2014;8(2):293–302. Ren F and Markovic D. A configurable 12-237 kS/s 12.8 mW sparseapproximation engine for mobile data aggregation of compressively sampled physiological signals. IEEE J Solid-State Circuits. 2015;PP(99):1–11. Calderbank R and Jafarpour S. Reed Muller sensing matrices and the LASSO. In: Carlet C, Pott A, editors. Sequences and Their Applications - SETA 2010. Berlin, Germany: Springer Berlin Heidelberg; 2010. pp. 442–463. Jafarpour S. Deterministic compressed sensing. Dept. Comput. Sci., Princeton Univ., Princeton, NJ; 2011. Hegde C, Sankaranarayanan AC, Yin W and Baraniuk RG. NuMax: A convex approach for learning near-isometric linear embeddings. IEEE Trans Signal Process. 2015;63(22):6109–6121. Chen F, Chandrakasan AP and Stojanovic VM. Design and analysis of a hardware-efficient compressed sensing architecture for data compression in wireless sensors. IEEE J Solid-State Circuits. 2012;47(3): 744–756. Kim S and Choi YK. Resistive switching of aluminum oxide for flexible memory. Appl Phys Lett. 2008;92(22):223508.1–223508.3.
236 [292]
[293]
[294]
[295]
[296]
[297] [298]
[299]
[300]
[301]
[302] [303]
ReRAM-based machine learning Baraniuk R, Davenport M, DeVore R and Wakin M. A simple proof of the restricted isometry property for random matrices. Constr Approx. 2008;28(3):253–263. Lee HY, Chen PS, Wu TY, et al. Low power and high speed bipolar switching with a thin reactive Ti buffer layer in robust HfO2 based RRAM. In: IEEE Electron Devices Meeting; 2008. Sheu S, Chiang P, Lin W, et al. A 5ns fast write multi-level non-volatile 1 K bits RRAM memory with advance write scheme. In: Proc. 2009 Symp. VLSI Circuits, Kyoto, Japan; 2009. pp. 82–83. Costa L and Oliveira P. Evolutionary algorithms approach to the solution of mixed integer non-linear programming problems. Comput Chem Eng. 2001;25(2):257–266. ApS M. The MOSEK optimization toolbox for MATLAB manual. Version 7.1 (Revision 28); 2015. Available from: http://docs.mosek.com/7.1/toolbox/ index.html. Sahinidis N. BARON: A general purpose global optimization software package. J Global Optim. 1996;8(2):201–205. Huang GB, Ramesh M, Berg T and Learned-Miller E. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Univ. of Massachusetts, Amherst, MA; 2007. 07-49. Moody GB and Mark RG. The impact of the MIT-BIH arrhythmia database. IEEE Eng Med Biol. 2001;20(3):45–50. Available from: http://www. physionet.org/physiobank/database/mitdb/. Wilton SJ and Jouppi NP. CACTI: An enhanced cache access and cycle time model. IEEE J Solid-State Circuits. 1996;31(5):677–688. Available from: http://www.hpl.hp.com/research/cacti/. ChenYS, Lee HY, Chen PS, et al. Highly scalable hafnium oxide memory with improvements of resistive distribution and read disturb immunity. In: Proc. 2009 IEEE Int. Electron Devices Meeting, Baltimore, MD; 2009. pp. 1–4. Wong HP, Lee H, Yu S, et al. Metal-oxide RRAM. Proc IEEE. 2012;100(6):1951–1970. Bavikadi S, Sutradhar PR, Khasawneh KN, Ganguly A and Sai Manoj PD. A review of in-memory computing architectures for machine learning applications. In: ACM Great Lakes Symposium on VLSI (GLSVLSI); 2020.
Index
accuracy analysis 182 accuracy under approximation 184–185 accuracy under device variation 182–183 peak accuracy comparison 182 Alpha Go 20 analog-to-digital converter (ADC) 30, 55, 99, 168 AND gate 70–71 artificial neural network (ANN) 20, 216 auto-encoder technique 97 auxiliary processing unit (APU) 39 backpropagation technique 81 BARON 202 batch normalization (BN) 168 Bernoulli matrix 190 big-data analytics 4 bijective function 145 binarization (Binrz) 90 binarized neural networks (BNNs) 132, 167 binary convolutional (BinConv) layers 132 binary convolutional neural network (BCNN) 16, 39, 88 accelerator 38 bitwise batch normalization 88–90 bitwise CNN model overview 90–93 bitwise convolution 88 bitwise pooling and activation functions 90 inference acceleration on 1S1R array
mapping batch normalization, pooling and binarization 122 mapping unsigned bitwise convolution 121–122 inference acceleration on passive array experimental settings 132 mapping bitwise batch normalization 118–120 mapping bitwise convolution 117–118 mapping bitwise pooling and binarization 120 performance evaluation 133 summary of mapping bitwise CNN 120–121 binary weighted neural network (BWNN) 36 biomedical wireless circuits 189 bit-width configuration analysis 151–152, 158 bitwise XNOR operation 88 central processing unit (CPU) 19, 168, 216 Cholesky decomposition 83 CMOS layer accelerator 145 layer implementation for decoding and incremental least-squares 143–145 CMOS-ASIC 216 CMP-PIM 40 compressive sensing Boolean embedding circuit CMOS-based 194–195
238
ReRAM-based machine learning
problem formulation 197 ReRAM crossbar-based 195–196 IH algorithm orthogonal rotation 198 overall optimization algorithm 199–200 quantization 199 and isometric distortion 192 numerical results experiment setup 203–204 hardware performance evaluation 210–213 IH algorithm on high-D ECG signals 204–207 row generation algorithm on low-D image patches 207–210 optimized near-isometric embedding 193–194 row generation algorithm convex relaxation of orthogonal constraint 201–202 elimination of norm equality constraint 200–201 overall optimization algorithm 202–203 computational unit-cum-memory 74–75 conventional communication protocol 110 convolutional neural network (CNN) 20, 87–88, 117, 167 binary 39 bitwise CNN 138 bitwise-parallelized 90–91 conventional CNN 138 direct-truncated 92–93 traditional direct mapping of 88 cores, pPIM 36 coupled ReRAM oscillator network 106–108 for L2-norm calculation 106–107 performance evaluation 107–108 data quantization 142 data storage
multi-threshold resistance for 62–63 ReRAM 61 deep convolutional neural networks (DCNNs) 87 binary convolutional neural network 88 convolutional neural network 87–88 deep learning for multilayer neural network 87 deep neural network (DNN) 27 with quantization 168 basics of 168–170 quantized activation function and pooling 172–173 quantized BN 172 quantized convolution and residual block 170–172 quantized deep neural network overview 173 training strategy 173–174 degree of- match (DOM) 125 density matrix 98 diffusive-type ReRAM device 52–54 digitalized ReRAM crossbar 55–57 digital processing unit (DPU) 39 digital-to-analog converter (DAC) 30, 55, 168 digitized computing-in-memory fashion 28 Discrete Fourier transform (DFT) analysis 60 distributed crossbar in-memory architecture communication protocol and control bus 110–111 memory-computing integration 109–110 distributed ReRAM-crossbar in-memory architecture 109 DrAcc 36 drift-type ReRAM device 45–52 DRISA 36 dynamic random access memory (DRAM) 5, 7–9, 19
Index encoding 67, 102, 116–117, 143 nonuniform 69–70 uniform input 68–69 energy-delay-product (EDP) 105, 151 energy efficiency 141 error correction code (ECC) techniques 34 feature extraction 82 first-come-first-serve (FCFS) 111 flash memory NAND 10 NOR 10 floating gate (FG) transistors 9 gate recurrent unit (GRU) 27 Gaussian distribution 133, 140, 160, 183 giga operations per second (GOPS) 137 graphics processing unit (GPU) 19, 127, 168 Group 36 Hadoop 3, 4 hard disk drive (HDD) 5 high-resistance state (HRS) 21, 196 ImageNet 167 IMCE 39–40 information technology 3 in-memory computing (IMC) architectures 3, 25 analog and digitized fashion of 27–28 analog MAC 29–30 analysis of 40–43 bitcell and array design of analog 31–32 cascading macros 30–31 challenges of analog 33–34 digitized IMC 34 DRAM-based 34–37 emerging memory technologies 40 ideal logic 11 at memory-block level 13
239
at memory-cell level 12 NAND-flash-based 37 nonvolatile 11–13 operating principles of 26–27 peripheral circuitry of analog 32–33 ReRAM-based 38–39 SOT-MRAM-based 39–40 SRAM-based 37 STT-MRAM-based 39 trade-offs of analog devices 34 traditional and 40 XIMA 15 in-pair control bus 109 internet-of-things (IoT) 3 ISAAC 38 iterative heuristic (IH) algorithm 207 on high-D ECG signals 204 orthogonal rotation 198 overall optimization algorithm 199–200 quantization 199 iterative row generation algorithm 208 Joglekar window function 49–50 least-square method 16 least squares (LS) solution 81 logic-in-memory architecture 11 long short-term memory (LSTM) 27 look-up table (LUT) 114 low-dimensional embedding 189 low-resistance state (LRS) 21 machine learning (ML) algorithms 79 adoption of 3 BCNN-based inference acceleration on passive array 117 deep convolutional neural network binary convolutional neural network 88 convolutional neural network 87–88 deep learning for multilayer neural network 87
240
ReRAM-based machine learning
experimental evaluation of 126, 151 IMC architectures: see in-memory computing (IMC) architectures L2-norm-gradient-based learning direct-gradient-based L2-norm optimization 86 and inference acceleration 123 multilayer neural network 84–86 single-layer feedforward network feature extraction 82 incremental LS solver-based learning 83–84 neural network-based learning 82–83 SLFN-based learning and inference acceleration 115 support vector machines 79–80 tensorized neural network tensor-train-based neural network 94–96 tensor-train decomposition and compression 93–94 training TNN 96–98 XOR operation for 101–102 magnetic tunneling junction (MTJ) 22 mapping bitwise batch normalization 118–120 bitwise CNN 120–121 bitwise convolution 117–118 bitwise pooling and binarization 120 matrix–vector multiplication 124 on multilayer architecture 145–149 MapReduce 3 matrix–vector multiplication 142–143 memory-computing integration 109–110 memory wall 3–6, 19 memristor device 61 mixed-integer nonlinear programming (MINLP) 200 mode size 93 modified alternating LS (MALS) method 93 Moore’s law 4 MOSEK 202
MOSFET transistor 9 most significant bit (MSB) 68 MRAM-DIMA 39 MRIMA 39 multilayer perceptron (MLP) 27 multilevel sense amplifiers (MSLAs) 39 multiplication-and-accumulation (MAC) 27, 192 nanoscale limitations 10 neural network (NN) 82–83 backpropagation training for 86 binary convolutional 88 biological 19 convolutional 87–88 deep learning for multilayer 87 energy-efficient tensor-compressed 160–162 layer-wise training of 97 multilayer 84–86 single-layer 81 tensor-train-based 94–96 neuromorphic computing 19 devices phase change memory 23–24 resistive random-access memory 21–22 spin-transfer-torque magnetic random-access memory 22 NVM devices linear/symmetric tuning 24 long time retention 25 low stochastic tunneling 25 multilevel resistance state 24–25 scalability and stackability 24 system on chip 20 non-distributed in-memory computing 103 nonuniform encoding 69–70 nonvolatile memories (NVMs) 4–5, 20, 167 characteristics of 24–25 flash 9–10 NuMax 190
Index Nvidia GeForce GTX 970 151 Nyquist sampling theorem 189 one-transistor-one-ReRAM (1T1R) 57 Opamp and feedback resistor 29 optical disk drive (ODD) 5 OR gate 70 orthogonal matching pursuit (OMP) technique 82 orthogonal transformation matrix 199
parallel digitizing 100–101, 115–116, 143 passive ReRAM array 57 performance analysis 159–160, 185 energy and area 185 throughput and efficiency 185–187 performance evaluation 156 phase-change memory (PCM) 20, 23–24 pipelayer 38 pooling 90, 172–173 power consumption 140–141 powerwall 3–6 principal component analysis (PCA) 127 processing-in-memory (PIM) 25 programmable PIM (pPIM) 36 programmable read-only memory (PROM) 5 pseudo-random number generator (PRNG) 195 puffer transformation 189 pulse-width-modulated (PWM) 39 QBN 172 QConv 171, 180 QPooling 173, 180 QReLUs 172–173, 180 quantization 199 Quantized BN 172 see also QBN quantum dynamics 98 quantum tunneling 9
241
read operation 7–8 readout circuit 74 readout method 64–65 reconstructed signal-to-noise ratio (RSNR) 191 rectified linear unit (ReLU) layers 168 recurrent neural network (RNN) 20 Reed–Muller code 189 refresh 8 residual network (ResNet) 167 accuracy analysis 182 accuracy under approximation 184–185 accuracy under device variation 182–183 peak accuracy comparison 182 deep neural network with quantization basics of 168–170 quantized activation function and pooling 172–173 quantized BN 172 quantized convolution and residual block 170–172 quantized deep neural network overview 173 training strategy 173–174 device simulations 181 experiment settings 180 performance analysis 185 energy and area 185 throughput and efficiency 185–187 on ReRAM crossbar 177 mapping strategy 177–178 overall architecture 178–180 resistive random access memory (ReRAM) 3, 19, 83, 99, 167, 190, 215 3D XIMA 3D multilayer CMOS-ReRAM architecture 112–114 3D single-layer CMOS-ReRAM architecture 111–112 area and energy consumption of 57
242
ReRAM-based machine learning
based inner-product operation 102 based oscillator 59–60 characteristics of 45 coupled oscillator network 106–108 crossbar 174–177 mapping strategy 177–178 overall architecture 178–180 crossbar network 99–105 mapping matrix–vector multiplication 124 crossbar structure 54 digitalized 55–57 direct-connected 57–58 one-selector-one 59 one-transistor-one 58–59 traditional analog 55 customized DAC and ADC circuits 176 data storage 61 device simulation 138–140 diffusive-type 52–54 distributed crossbar in-memory architecture communication protocol and control bus 110–111 memory-computing integration 109–110 drift-type 45–52 encoding and three-bit storage 67 exploration of the memristance range 67–68 nonuniform encoding 69–70 uniform input encoding 68–69 in-memory computing architecture 176–177 layer accelerator 145 logic functional units AND gate 70–71 OR gate 70 for logic operations based circuits 73–74 as computational unit-cum-memory 74–75 simulation settings 72–73 mathematic model of 22
multi-threshold resistance for data storage 62–63 and nonlinear effects for dynamic model 47 nonvolatile 13 one-selector-one 15 one-transistor-one 15 operating principle of 46 oscillator network mapping L2-norm calculation on coupled 124–125 spiking neural networks on 216 validation 65–67 write and read 63 ResNet with 50 layers (ResNet-50) 168 restricted isometry property (RIP) 192 row generation algorithm convex relaxation of orthogonal constraint 201–202 elimination of norm equality constraint 200–201 on low-D image patches 207 overall optimization algorithm 202–203 scalability analysis 151 SCOPE 36, 40, 42 Semiconductor memory memory technologies 5–10 nanoscale limitations 10 Sigmoid function 83, 97, 144 single-layer feedforward network (SLFN) 80–81 feature extraction 82 incremental LS solver-based learning 83–84 neural network-based learning 82–83 single-layer feedforward neural network (SLFN) 16 learning and inference acceleration adding and shifting for inner-product result 117 encoding 116–117 experiment settings 127
Index parallel digitizing 115–116 performance evaluation 127–129 XOR 116 learning and inference procedures of 216 on-chip design for 141–142 SPICE model 45 spike timing-dependent plasticity (STDP) 14 spiking neural network (SNN) 20 spin-torque-transfer-magnetic random-access memory (STT-MRAM) 20, 22 static random access memories (SRAMs) 5–7, 21 support vector decomposition (SVD) 98, 155 support vector machine (SVM) 16, 79–80, 151 based machine learning 79–80 binary classifier 79 fingerprint recognition by multiple 80 tensorized neural network (TNN) 16, 112, 141 on 3D CMOS-ReRAM architecture 149–150 distributed on-chip design on 3D single-layer architecture 160–163 layer-wise training of 93 on-chip design for 145 on-chip design on 3D multilayer architecture 155–160 tensor-train-based neural network 94–96
243
tensor-train decomposition and compression 93–94 training 96–98 tensors train-based neural network 94–96 train decomposition and compression 93–94 ternary weighted neural network (TWNN) 36 three-dimensional (3D) CMOS-ReRAM architecture 215 three-dimensional (3D) multilayer CMOS-ReRAM architecture 112–114 three-dimensional (3D) single-layer CMOS-ReRAM architecture 111–112 three-dimensional (3D) XIMA 3D multilayer CMOS-ReRAM architecture 112–114 3D single-layer CMOS-ReRAM architecture 111–112 machine learning algorithms 141 traditional analog ReRAM crossbar 55 trans-impedance amplifier (TIA) 29 transition metal oxides (TMOs) 54 two-dimensional (2D) arrays 93 uniform input encoding 68–69 validation 65–67 Von Neumann architecture 20 write-in method 64 XNOR-ReRAM 38–39 XOR 101–102, 116, 143